{ "schema_version": 1, "status": "FINAL_REVIEWED_V1", "generated_at_utc": "2026-10-08T17:54:17.626733+00:00", "all_scheduled_final_texts_reviewed": true, "first_attempt_only": true, "browser_observations_included": false, "final_run_snapshot": true, "run_stop_reason": null, "evaluation_date_utc": "2026-10-08", "frozen_at_utc": "2026-10-08T17:17:52.794650+00:00", "commitment_sha256": "b1a3c86878e70bd7d6bb7fdf3aa3bce44826b5ab059ab734d281073df28f104c", "verified_frozen_analysis_inputs": [ { "file": "schedule.json", "sha256": "5e8a124e88b387077231125549d9e15e7d7eecb4afd95d165fe4e34d6557c14f" }, { "file": "private/api-cohort.json", "sha256": "c52d0e39350fc238363d20e43bdc117a4f248ac3cfe4478adbe08940d439fdef" }, { "file": "methods/api-response-schema.json", "sha256": "4732796378916e737503a48a7fab86ee76c7b5ea7a6218519a1d5170ddd4e9c4" }, { "file": "methods/api-scoring-plan.json", "sha256": "9ee65d7d00dc3c12136eaf338782f30a241f6246265c24423665fd6a754bbfd1" } ], "review_provenance": [ { "file": "private/api-review-M.json", "sha256": "2b0479986cdbefce5210137958fab58ad16d12deb006e2d2009375cb4d262191" }, { "file": "private/api-review-W.json", "sha256": "b1a0328a157c093e343eda0efda4d43bf94cfd61c34709910981c791c865619c" }, { "file": "private/api-review-E.json", "sha256": "15b180590ca16252d3673f237c6e9362348d74e05889f2403f94eb92b7ab1cb5" } ], "ignored_unscheduled_api_case_artifacts": [], "selected_company_count": 96, "selected_company_cohorts": { "removed": 64, "current_control": 32 }, "selected_company_strata": { "removed-20260528": 9, "removed-20260727": 27, "removed-20261005": 28, "listed-control-SDNT": 20, "listed-control-IRAQ2": 8, "listed-control-SDNTK": 4 }, "totals": { "scheduled_requests": 1152, "selected_companies": 96, "primary_eligible_requests": 1152, "artifacts_present": 1152, "complete_requests": 1152, "operational_counts": { "complete": 1152, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 211, "correct_absent_claim": 292, "wrong_present_claim": 29, "wrong_absent_claim": 0, "unresolved": 620, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 503, "wrong_requests": 29, "uncertain_requests": 620, "full_text_reviewed_requests": 1152, "web_calls_observed": 1606, "completed_web_actions_observed": 1362, "uncompleted_web_items_observed": 244, "requests_with_web_calls": 384, "requests_with_validation_flags": 533 }, "by_model_arm_cohort": [ { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "scheduled_requests": 64, "selected_companies": 32, "primary_eligible_requests": 64, "artifacts_present": 64, "complete_requests": 64, "operational_counts": { "complete": 64, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 64, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 64, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 64, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "scheduled_requests": 128, "selected_companies": 64, "primary_eligible_requests": 128, "artifacts_present": 128, "complete_requests": 128, "operational_counts": { "complete": 128, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 126, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 2, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 126, "wrong_requests": 0, "uncertain_requests": 2, "full_text_reviewed_requests": 128, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "scheduled_requests": 64, "selected_companies": 32, "primary_eligible_requests": 64, "artifacts_present": 64, "complete_requests": 64, "operational_counts": { "complete": 64, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 64, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 64, "full_text_reviewed_requests": 64, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 27 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "scheduled_requests": 128, "selected_companies": 64, "primary_eligible_requests": 128, "artifacts_present": 128, "complete_requests": 128, "operational_counts": { "complete": 128, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 128, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 128, "full_text_reviewed_requests": 128, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 61 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "scheduled_requests": 64, "selected_companies": 32, "primary_eligible_requests": 64, "artifacts_present": 64, "complete_requests": 64, "operational_counts": { "complete": 64, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 32, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 32, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 32, "wrong_requests": 0, "uncertain_requests": 32, "full_text_reviewed_requests": 64, "web_calls_observed": 195, "completed_web_actions_observed": 174, "uncompleted_web_items_observed": 21, "requests_with_web_calls": 64, "requests_with_validation_flags": 33 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "scheduled_requests": 128, "selected_companies": 64, "primary_eligible_requests": 128, "artifacts_present": 128, "complete_requests": 128, "operational_counts": { "complete": 128, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 11, "wrong_present_claim": 28, "wrong_absent_claim": 0, "unresolved": 89, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 11, "wrong_requests": 28, "uncertain_requests": 89, "full_text_reviewed_requests": 128, "web_calls_observed": 590, "completed_web_actions_observed": 488, "uncompleted_web_items_observed": 102, "requests_with_web_calls": 128, "requests_with_validation_flags": 113 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "scheduled_requests": 64, "selected_companies": 32, "primary_eligible_requests": 64, "artifacts_present": 64, "complete_requests": 64, "operational_counts": { "complete": 64, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 64, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 64, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 64, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "scheduled_requests": 128, "selected_companies": 64, "primary_eligible_requests": 128, "artifacts_present": 128, "complete_requests": 128, "operational_counts": { "complete": 128, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 128, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 128, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 128, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 5 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "scheduled_requests": 64, "selected_companies": 32, "primary_eligible_requests": 64, "artifacts_present": 64, "complete_requests": 64, "operational_counts": { "complete": 64, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 64, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 64, "full_text_reviewed_requests": 64, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 55 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "scheduled_requests": 128, "selected_companies": 64, "primary_eligible_requests": 128, "artifacts_present": 128, "complete_requests": 128, "operational_counts": { "complete": 128, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 128, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 128, "full_text_reviewed_requests": 128, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 109 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "scheduled_requests": 64, "selected_companies": 32, "primary_eligible_requests": 64, "artifacts_present": 64, "complete_requests": 64, "operational_counts": { "complete": 64, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 51, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 13, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 51, "wrong_requests": 0, "uncertain_requests": 13, "full_text_reviewed_requests": 64, "web_calls_observed": 204, "completed_web_actions_observed": 190, "uncompleted_web_items_observed": 14, "requests_with_web_calls": 64, "requests_with_validation_flags": 14 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "scheduled_requests": 128, "selected_companies": 64, "primary_eligible_requests": 128, "artifacts_present": 128, "complete_requests": 128, "operational_counts": { "complete": 128, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 27, "wrong_present_claim": 1, "wrong_absent_claim": 0, "unresolved": 100, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 27, "wrong_requests": 1, "uncertain_requests": 100, "full_text_reviewed_requests": 128, "web_calls_observed": 617, "completed_web_actions_observed": 510, "uncompleted_web_items_observed": 107, "requests_with_web_calls": 128, "requests_with_validation_flags": 116 } ], "by_model_arm_cohort_event_stratum": [ { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "scheduled_requests": 16, "selected_companies": 8, "primary_eligible_requests": 16, "artifacts_present": 16, "complete_requests": 16, "operational_counts": { "complete": 16, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 16, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 16, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 16, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "scheduled_requests": 40, "selected_companies": 20, "primary_eligible_requests": 40, "artifacts_present": 40, "complete_requests": 40, "operational_counts": { "complete": 40, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 40, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 40, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 40, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "scheduled_requests": 8, "selected_companies": 4, "primary_eligible_requests": 8, "artifacts_present": 8, "complete_requests": 8, "operational_counts": { "complete": 8, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 8, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 8, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 8, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "scheduled_requests": 18, "selected_companies": 9, "primary_eligible_requests": 18, "artifacts_present": 18, "complete_requests": 18, "operational_counts": { "complete": 18, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 18, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 18, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 18, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "scheduled_requests": 54, "selected_companies": 27, "primary_eligible_requests": 54, "artifacts_present": 54, "complete_requests": 54, "operational_counts": { "complete": 54, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 52, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 2, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 52, "wrong_requests": 0, "uncertain_requests": 2, "full_text_reviewed_requests": 54, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "scheduled_requests": 56, "selected_companies": 28, "primary_eligible_requests": 56, "artifacts_present": 56, "complete_requests": 56, "operational_counts": { "complete": 56, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 56, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 56, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 56, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "scheduled_requests": 16, "selected_companies": 8, "primary_eligible_requests": 16, "artifacts_present": 16, "complete_requests": 16, "operational_counts": { "complete": 16, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 16, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 16, "full_text_reviewed_requests": 16, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 1 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "scheduled_requests": 40, "selected_companies": 20, "primary_eligible_requests": 40, "artifacts_present": 40, "complete_requests": 40, "operational_counts": { "complete": 40, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 40, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 40, "full_text_reviewed_requests": 40, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 24 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "scheduled_requests": 8, "selected_companies": 4, "primary_eligible_requests": 8, "artifacts_present": 8, "complete_requests": 8, "operational_counts": { "complete": 8, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 8, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 8, "full_text_reviewed_requests": 8, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 2 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "scheduled_requests": 18, "selected_companies": 9, "primary_eligible_requests": 18, "artifacts_present": 18, "complete_requests": 18, "operational_counts": { "complete": 18, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 18, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 18, "full_text_reviewed_requests": 18, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 15 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "scheduled_requests": 54, "selected_companies": 27, "primary_eligible_requests": 54, "artifacts_present": 54, "complete_requests": 54, "operational_counts": { "complete": 54, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 54, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 54, "full_text_reviewed_requests": 54, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 17 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "scheduled_requests": 56, "selected_companies": 28, "primary_eligible_requests": 56, "artifacts_present": 56, "complete_requests": 56, "operational_counts": { "complete": 56, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 56, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 56, "full_text_reviewed_requests": 56, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 29 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "scheduled_requests": 16, "selected_companies": 8, "primary_eligible_requests": 16, "artifacts_present": 16, "complete_requests": 16, "operational_counts": { "complete": 16, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 7, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 9, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 7, "wrong_requests": 0, "uncertain_requests": 9, "full_text_reviewed_requests": 16, "web_calls_observed": 49, "completed_web_actions_observed": 44, "uncompleted_web_items_observed": 5, "requests_with_web_calls": 16, "requests_with_validation_flags": 8 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "scheduled_requests": 40, "selected_companies": 20, "primary_eligible_requests": 40, "artifacts_present": 40, "complete_requests": 40, "operational_counts": { "complete": 40, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 22, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 18, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 22, "wrong_requests": 0, "uncertain_requests": 18, "full_text_reviewed_requests": 40, "web_calls_observed": 121, "completed_web_actions_observed": 107, "uncompleted_web_items_observed": 14, "requests_with_web_calls": 40, "requests_with_validation_flags": 22 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "scheduled_requests": 8, "selected_companies": 4, "primary_eligible_requests": 8, "artifacts_present": 8, "complete_requests": 8, "operational_counts": { "complete": 8, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 3, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 5, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 3, "wrong_requests": 0, "uncertain_requests": 5, "full_text_reviewed_requests": 8, "web_calls_observed": 25, "completed_web_actions_observed": 23, "uncompleted_web_items_observed": 2, "requests_with_web_calls": 8, "requests_with_validation_flags": 3 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "scheduled_requests": 18, "selected_companies": 9, "primary_eligible_requests": 18, "artifacts_present": 18, "complete_requests": 18, "operational_counts": { "complete": 18, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 18, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 18, "full_text_reviewed_requests": 18, "web_calls_observed": 81, "completed_web_actions_observed": 69, "uncompleted_web_items_observed": 12, "requests_with_web_calls": 18, "requests_with_validation_flags": 15 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "scheduled_requests": 54, "selected_companies": 27, "primary_eligible_requests": 54, "artifacts_present": 54, "complete_requests": 54, "operational_counts": { "complete": 54, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 9, "wrong_absent_claim": 0, "unresolved": 45, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 9, "uncertain_requests": 45, "full_text_reviewed_requests": 54, "web_calls_observed": 262, "completed_web_actions_observed": 213, "uncompleted_web_items_observed": 49, "requests_with_web_calls": 54, "requests_with_validation_flags": 51 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "scheduled_requests": 56, "selected_companies": 28, "primary_eligible_requests": 56, "artifacts_present": 56, "complete_requests": 56, "operational_counts": { "complete": 56, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 11, "wrong_present_claim": 19, "wrong_absent_claim": 0, "unresolved": 26, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 11, "wrong_requests": 19, "uncertain_requests": 26, "full_text_reviewed_requests": 56, "web_calls_observed": 247, "completed_web_actions_observed": 206, "uncompleted_web_items_observed": 41, "requests_with_web_calls": 56, "requests_with_validation_flags": 47 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "scheduled_requests": 16, "selected_companies": 8, "primary_eligible_requests": 16, "artifacts_present": 16, "complete_requests": 16, "operational_counts": { "complete": 16, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 16, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 16, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 16, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "scheduled_requests": 40, "selected_companies": 20, "primary_eligible_requests": 40, "artifacts_present": 40, "complete_requests": 40, "operational_counts": { "complete": 40, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 40, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 40, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 40, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "scheduled_requests": 8, "selected_companies": 4, "primary_eligible_requests": 8, "artifacts_present": 8, "complete_requests": 8, "operational_counts": { "complete": 8, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 8, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 8, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 8, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "scheduled_requests": 18, "selected_companies": 9, "primary_eligible_requests": 18, "artifacts_present": 18, "complete_requests": 18, "operational_counts": { "complete": 18, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 18, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 18, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 18, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "scheduled_requests": 54, "selected_companies": 27, "primary_eligible_requests": 54, "artifacts_present": 54, "complete_requests": 54, "operational_counts": { "complete": 54, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 54, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 54, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 54, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 3 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "scheduled_requests": 56, "selected_companies": 28, "primary_eligible_requests": 56, "artifacts_present": 56, "complete_requests": 56, "operational_counts": { "complete": 56, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 56, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 56, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 56, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 2 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "scheduled_requests": 16, "selected_companies": 8, "primary_eligible_requests": 16, "artifacts_present": 16, "complete_requests": 16, "operational_counts": { "complete": 16, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 16, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 16, "full_text_reviewed_requests": 16, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 9 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "scheduled_requests": 40, "selected_companies": 20, "primary_eligible_requests": 40, "artifacts_present": 40, "complete_requests": 40, "operational_counts": { "complete": 40, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 40, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 40, "full_text_reviewed_requests": 40, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 40 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "scheduled_requests": 8, "selected_companies": 4, "primary_eligible_requests": 8, "artifacts_present": 8, "complete_requests": 8, "operational_counts": { "complete": 8, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 8, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 8, "full_text_reviewed_requests": 8, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 6 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "scheduled_requests": 18, "selected_companies": 9, "primary_eligible_requests": 18, "artifacts_present": 18, "complete_requests": 18, "operational_counts": { "complete": 18, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 18, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 18, "full_text_reviewed_requests": 18, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 18 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "scheduled_requests": 54, "selected_companies": 27, "primary_eligible_requests": 54, "artifacts_present": 54, "complete_requests": 54, "operational_counts": { "complete": 54, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 54, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 54, "full_text_reviewed_requests": 54, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 38 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "scheduled_requests": 56, "selected_companies": 28, "primary_eligible_requests": 56, "artifacts_present": 56, "complete_requests": 56, "operational_counts": { "complete": 56, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 56, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 56, "full_text_reviewed_requests": 56, "web_calls_observed": 0, "completed_web_actions_observed": 0, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 0, "requests_with_validation_flags": 53 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "scheduled_requests": 16, "selected_companies": 8, "primary_eligible_requests": 16, "artifacts_present": 16, "complete_requests": 16, "operational_counts": { "complete": 16, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 14, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 2, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 14, "wrong_requests": 0, "uncertain_requests": 2, "full_text_reviewed_requests": 16, "web_calls_observed": 40, "completed_web_actions_observed": 38, "uncompleted_web_items_observed": 2, "requests_with_web_calls": 16, "requests_with_validation_flags": 2 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "scheduled_requests": 40, "selected_companies": 20, "primary_eligible_requests": 40, "artifacts_present": 40, "complete_requests": 40, "operational_counts": { "complete": 40, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 29, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 11, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 29, "wrong_requests": 0, "uncertain_requests": 11, "full_text_reviewed_requests": 40, "web_calls_observed": 140, "completed_web_actions_observed": 128, "uncompleted_web_items_observed": 12, "requests_with_web_calls": 40, "requests_with_validation_flags": 12 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "scheduled_requests": 8, "selected_companies": 4, "primary_eligible_requests": 8, "artifacts_present": 8, "complete_requests": 8, "operational_counts": { "complete": 8, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 8, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 0, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 8, "wrong_requests": 0, "uncertain_requests": 0, "full_text_reviewed_requests": 8, "web_calls_observed": 24, "completed_web_actions_observed": 24, "uncompleted_web_items_observed": 0, "requests_with_web_calls": 8, "requests_with_validation_flags": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "scheduled_requests": 18, "selected_companies": 9, "primary_eligible_requests": 18, "artifacts_present": 18, "complete_requests": 18, "operational_counts": { "complete": 18, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 0, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 18, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 0, "wrong_requests": 0, "uncertain_requests": 18, "full_text_reviewed_requests": 18, "web_calls_observed": 85, "completed_web_actions_observed": 72, "uncompleted_web_items_observed": 13, "requests_with_web_calls": 18, "requests_with_validation_flags": 15 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "scheduled_requests": 54, "selected_companies": 27, "primary_eligible_requests": 54, "artifacts_present": 54, "complete_requests": 54, "operational_counts": { "complete": 54, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 1, "wrong_present_claim": 0, "wrong_absent_claim": 0, "unresolved": 53, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 1, "wrong_requests": 0, "uncertain_requests": 53, "full_text_reviewed_requests": 54, "web_calls_observed": 257, "completed_web_actions_observed": 216, "uncompleted_web_items_observed": 41, "requests_with_web_calls": 54, "requests_with_validation_flags": 46 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "scheduled_requests": 56, "selected_companies": 28, "primary_eligible_requests": 56, "artifacts_present": 56, "complete_requests": 56, "operational_counts": { "complete": 56, "transport_failure": 0, "rate_limit": 0, "provider_error": 0, "schema_invalid": 0, "length_terminated": 0, "ambiguous_delivery": 0, "not_dispatched_budget": 0, "source_temporally_contaminated": 0, "pending_not_dispatched": 0, "pending_in_flight": 0, "not_dispatched_stopped": 0, "artifact_integrity_failure": 0 }, "membership_counts": { "correct_present_claim": 0, "correct_absent_claim": 26, "wrong_present_claim": 1, "wrong_absent_claim": 0, "unresolved": 29, "contradictory": 0, "out_of_scope": 0, "no_usable_answer": 0 }, "correct_requests": 26, "wrong_requests": 1, "uncertain_requests": 29, "full_text_reviewed_requests": 56, "web_calls_observed": 275, "completed_web_actions_observed": 222, "uncompleted_web_items_observed": 53, "requests_with_web_calls": 56, "requests_with_validation_flags": 55 } ], "company_two_repetition_outcomes": [ { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-001-gpt-6-luna-E-r1", "api-case-001-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-002-gpt-6-luna-E-r1", "api-case-002-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-003-gpt-6-luna-E-r1", "api-case-003-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-004-gpt-6-luna-E-r1", "api-case-004-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-005-gpt-6-luna-E-r1", "api-case-005-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-006-gpt-6-luna-E-r1", "api-case-006-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-007-gpt-6-luna-E-r1", "api-case-007-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-008-gpt-6-luna-E-r1", "api-case-008-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-009-gpt-6-luna-E-r1", "api-case-009-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-010-gpt-6-luna-E-r1", "api-case-010-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-011-gpt-6-luna-E-r1", "api-case-011-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-012-gpt-6-luna-E-r1", "api-case-012-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-013-gpt-6-luna-E-r1", "api-case-013-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-014-gpt-6-luna-E-r1", "api-case-014-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-015-gpt-6-luna-E-r1", "api-case-015-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-016-gpt-6-luna-E-r1", "api-case-016-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-017-gpt-6-luna-E-r1", "api-case-017-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-018-gpt-6-luna-E-r1", "api-case-018-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-019-gpt-6-luna-E-r1", "api-case-019-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-020-gpt-6-luna-E-r1", "api-case-020-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-021-gpt-6-luna-E-r1", "api-case-021-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-022-gpt-6-luna-E-r1", "api-case-022-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-023-gpt-6-luna-E-r1", "api-case-023-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-024-gpt-6-luna-E-r1", "api-case-024-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-025-gpt-6-luna-E-r1", "api-case-025-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-026-gpt-6-luna-E-r1", "api-case-026-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-027-gpt-6-luna-E-r1", "api-case-027-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-028-gpt-6-luna-E-r1", "api-case-028-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-029-gpt-6-luna-E-r1", "api-case-029-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-030-gpt-6-luna-E-r1", "api-case-030-gpt-6-luna-E-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-031-gpt-6-luna-E-r1", "api-case-031-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-032-gpt-6-luna-E-r1", "api-case-032-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-033-gpt-6-luna-E-r1", "api-case-033-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-034-gpt-6-luna-E-r1", "api-case-034-gpt-6-luna-E-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-035-gpt-6-luna-E-r1", "api-case-035-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-036-gpt-6-luna-E-r1", "api-case-036-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-037-gpt-6-luna-E-r1", "api-case-037-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-038-gpt-6-luna-E-r1", "api-case-038-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-039-gpt-6-luna-E-r1", "api-case-039-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-040-gpt-6-luna-E-r1", "api-case-040-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-041-gpt-6-luna-E-r1", "api-case-041-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-042-gpt-6-luna-E-r1", "api-case-042-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-043-gpt-6-luna-E-r1", "api-case-043-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-044-gpt-6-luna-E-r1", "api-case-044-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-045-gpt-6-luna-E-r1", "api-case-045-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-046-gpt-6-luna-E-r1", "api-case-046-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-047-gpt-6-luna-E-r1", "api-case-047-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-048-gpt-6-luna-E-r1", "api-case-048-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-049-gpt-6-luna-E-r1", "api-case-049-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-050-gpt-6-luna-E-r1", "api-case-050-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-051-gpt-6-luna-E-r1", "api-case-051-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-052-gpt-6-luna-E-r1", "api-case-052-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-053-gpt-6-luna-E-r1", "api-case-053-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-054-gpt-6-luna-E-r1", "api-case-054-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-055-gpt-6-luna-E-r1", "api-case-055-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-056-gpt-6-luna-E-r1", "api-case-056-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-057-gpt-6-luna-E-r1", "api-case-057-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-058-gpt-6-luna-E-r1", "api-case-058-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-059-gpt-6-luna-E-r1", "api-case-059-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-060-gpt-6-luna-E-r1", "api-case-060-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-061-gpt-6-luna-E-r1", "api-case-061-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-062-gpt-6-luna-E-r1", "api-case-062-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-063-gpt-6-luna-E-r1", "api-case-063-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-064-gpt-6-luna-E-r1", "api-case-064-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-065-gpt-6-luna-E-r1", "api-case-065-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-066-gpt-6-luna-E-r1", "api-case-066-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-067-gpt-6-luna-E-r1", "api-case-067-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-068-gpt-6-luna-E-r1", "api-case-068-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-069-gpt-6-luna-E-r1", "api-case-069-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-070-gpt-6-luna-E-r1", "api-case-070-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-071-gpt-6-luna-E-r1", "api-case-071-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-072-gpt-6-luna-E-r1", "api-case-072-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-073-gpt-6-luna-E-r1", "api-case-073-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-074-gpt-6-luna-E-r1", "api-case-074-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-075-gpt-6-luna-E-r1", "api-case-075-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-076-gpt-6-luna-E-r1", "api-case-076-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-077-gpt-6-luna-E-r1", "api-case-077-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-078-gpt-6-luna-E-r1", "api-case-078-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-079-gpt-6-luna-E-r1", "api-case-079-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-080-gpt-6-luna-E-r1", "api-case-080-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-081-gpt-6-luna-E-r1", "api-case-081-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-082-gpt-6-luna-E-r1", "api-case-082-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-083-gpt-6-luna-E-r1", "api-case-083-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-084-gpt-6-luna-E-r1", "api-case-084-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-085-gpt-6-luna-E-r1", "api-case-085-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-086-gpt-6-luna-E-r1", "api-case-086-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-087-gpt-6-luna-E-r1", "api-case-087-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-088-gpt-6-luna-E-r1", "api-case-088-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-089-gpt-6-luna-E-r1", "api-case-089-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-090-gpt-6-luna-E-r1", "api-case-090-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-091-gpt-6-luna-E-r1", "api-case-091-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-092-gpt-6-luna-E-r1", "api-case-092-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-093-gpt-6-luna-E-r1", "api-case-093-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-094-gpt-6-luna-E-r1", "api-case-094-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-095-gpt-6-luna-E-r1", "api-case-095-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "E", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-096-gpt-6-luna-E-r1", "api-case-096-gpt-6-luna-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-001-gpt-6-luna-M-r1", "api-case-001-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-002-gpt-6-luna-M-r1", "api-case-002-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-003-gpt-6-luna-M-r1", "api-case-003-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-004-gpt-6-luna-M-r1", "api-case-004-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-005-gpt-6-luna-M-r1", "api-case-005-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-006-gpt-6-luna-M-r1", "api-case-006-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-007-gpt-6-luna-M-r1", "api-case-007-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-008-gpt-6-luna-M-r1", "api-case-008-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-009-gpt-6-luna-M-r1", "api-case-009-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-010-gpt-6-luna-M-r1", "api-case-010-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-011-gpt-6-luna-M-r1", "api-case-011-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-012-gpt-6-luna-M-r1", "api-case-012-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-013-gpt-6-luna-M-r1", "api-case-013-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-014-gpt-6-luna-M-r1", "api-case-014-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-015-gpt-6-luna-M-r1", "api-case-015-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-016-gpt-6-luna-M-r1", "api-case-016-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-017-gpt-6-luna-M-r1", "api-case-017-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-018-gpt-6-luna-M-r1", "api-case-018-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-019-gpt-6-luna-M-r1", "api-case-019-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-020-gpt-6-luna-M-r1", "api-case-020-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-021-gpt-6-luna-M-r1", "api-case-021-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-022-gpt-6-luna-M-r1", "api-case-022-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-023-gpt-6-luna-M-r1", "api-case-023-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-024-gpt-6-luna-M-r1", "api-case-024-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-025-gpt-6-luna-M-r1", "api-case-025-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-026-gpt-6-luna-M-r1", "api-case-026-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-027-gpt-6-luna-M-r1", "api-case-027-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-028-gpt-6-luna-M-r1", "api-case-028-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-029-gpt-6-luna-M-r1", "api-case-029-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-030-gpt-6-luna-M-r1", "api-case-030-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-031-gpt-6-luna-M-r1", "api-case-031-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-032-gpt-6-luna-M-r1", "api-case-032-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-033-gpt-6-luna-M-r1", "api-case-033-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-034-gpt-6-luna-M-r1", "api-case-034-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-035-gpt-6-luna-M-r1", "api-case-035-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-036-gpt-6-luna-M-r1", "api-case-036-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-037-gpt-6-luna-M-r1", "api-case-037-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-038-gpt-6-luna-M-r1", "api-case-038-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-039-gpt-6-luna-M-r1", "api-case-039-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-040-gpt-6-luna-M-r1", "api-case-040-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-041-gpt-6-luna-M-r1", "api-case-041-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-042-gpt-6-luna-M-r1", "api-case-042-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-043-gpt-6-luna-M-r1", "api-case-043-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-044-gpt-6-luna-M-r1", "api-case-044-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-045-gpt-6-luna-M-r1", "api-case-045-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-046-gpt-6-luna-M-r1", "api-case-046-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-047-gpt-6-luna-M-r1", "api-case-047-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-048-gpt-6-luna-M-r1", "api-case-048-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-049-gpt-6-luna-M-r1", "api-case-049-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-050-gpt-6-luna-M-r1", "api-case-050-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-051-gpt-6-luna-M-r1", "api-case-051-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-052-gpt-6-luna-M-r1", "api-case-052-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-053-gpt-6-luna-M-r1", "api-case-053-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-054-gpt-6-luna-M-r1", "api-case-054-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-055-gpt-6-luna-M-r1", "api-case-055-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-056-gpt-6-luna-M-r1", "api-case-056-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-057-gpt-6-luna-M-r1", "api-case-057-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-058-gpt-6-luna-M-r1", "api-case-058-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-059-gpt-6-luna-M-r1", "api-case-059-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-060-gpt-6-luna-M-r1", "api-case-060-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-061-gpt-6-luna-M-r1", "api-case-061-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-062-gpt-6-luna-M-r1", "api-case-062-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-063-gpt-6-luna-M-r1", "api-case-063-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-064-gpt-6-luna-M-r1", "api-case-064-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-065-gpt-6-luna-M-r1", "api-case-065-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-066-gpt-6-luna-M-r1", "api-case-066-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-067-gpt-6-luna-M-r1", "api-case-067-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-068-gpt-6-luna-M-r1", "api-case-068-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-069-gpt-6-luna-M-r1", "api-case-069-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-070-gpt-6-luna-M-r1", "api-case-070-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-071-gpt-6-luna-M-r1", "api-case-071-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-072-gpt-6-luna-M-r1", "api-case-072-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-073-gpt-6-luna-M-r1", "api-case-073-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-074-gpt-6-luna-M-r1", "api-case-074-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-075-gpt-6-luna-M-r1", "api-case-075-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-076-gpt-6-luna-M-r1", "api-case-076-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-077-gpt-6-luna-M-r1", "api-case-077-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-078-gpt-6-luna-M-r1", "api-case-078-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-079-gpt-6-luna-M-r1", "api-case-079-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-080-gpt-6-luna-M-r1", "api-case-080-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-081-gpt-6-luna-M-r1", "api-case-081-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-082-gpt-6-luna-M-r1", "api-case-082-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-083-gpt-6-luna-M-r1", "api-case-083-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-084-gpt-6-luna-M-r1", "api-case-084-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-085-gpt-6-luna-M-r1", "api-case-085-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-086-gpt-6-luna-M-r1", "api-case-086-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-087-gpt-6-luna-M-r1", "api-case-087-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-088-gpt-6-luna-M-r1", "api-case-088-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-089-gpt-6-luna-M-r1", "api-case-089-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-090-gpt-6-luna-M-r1", "api-case-090-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-091-gpt-6-luna-M-r1", "api-case-091-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-092-gpt-6-luna-M-r1", "api-case-092-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-093-gpt-6-luna-M-r1", "api-case-093-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-094-gpt-6-luna-M-r1", "api-case-094-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-095-gpt-6-luna-M-r1", "api-case-095-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "M", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-096-gpt-6-luna-M-r1", "api-case-096-gpt-6-luna-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-001-gpt-6-luna-W-r1", "api-case-001-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-002-gpt-6-luna-W-r1", "api-case-002-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-003-gpt-6-luna-W-r1", "api-case-003-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-004-gpt-6-luna-W-r1", "api-case-004-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-005-gpt-6-luna-W-r1", "api-case-005-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-006-gpt-6-luna-W-r1", "api-case-006-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-007-gpt-6-luna-W-r1", "api-case-007-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-008-gpt-6-luna-W-r1", "api-case-008-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-009-gpt-6-luna-W-r1", "api-case-009-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-010-gpt-6-luna-W-r1", "api-case-010-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-011-gpt-6-luna-W-r1", "api-case-011-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-012-gpt-6-luna-W-r1", "api-case-012-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-013-gpt-6-luna-W-r1", "api-case-013-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 2, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 1.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-014-gpt-6-luna-W-r1", "api-case-014-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-015-gpt-6-luna-W-r1", "api-case-015-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-016-gpt-6-luna-W-r1", "api-case-016-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-017-gpt-6-luna-W-r1", "api-case-017-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-018-gpt-6-luna-W-r1", "api-case-018-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-019-gpt-6-luna-W-r1", "api-case-019-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-020-gpt-6-luna-W-r1", "api-case-020-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-021-gpt-6-luna-W-r1", "api-case-021-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-022-gpt-6-luna-W-r1", "api-case-022-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-023-gpt-6-luna-W-r1", "api-case-023-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-024-gpt-6-luna-W-r1", "api-case-024-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-025-gpt-6-luna-W-r1", "api-case-025-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-026-gpt-6-luna-W-r1", "api-case-026-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-027-gpt-6-luna-W-r1", "api-case-027-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-028-gpt-6-luna-W-r1", "api-case-028-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-029-gpt-6-luna-W-r1", "api-case-029-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-030-gpt-6-luna-W-r1", "api-case-030-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-031-gpt-6-luna-W-r1", "api-case-031-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-032-gpt-6-luna-W-r1", "api-case-032-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-033-gpt-6-luna-W-r1", "api-case-033-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-034-gpt-6-luna-W-r1", "api-case-034-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-035-gpt-6-luna-W-r1", "api-case-035-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-036-gpt-6-luna-W-r1", "api-case-036-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-037-gpt-6-luna-W-r1", "api-case-037-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 2, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 1.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-038-gpt-6-luna-W-r1", "api-case-038-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-039-gpt-6-luna-W-r1", "api-case-039-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-040-gpt-6-luna-W-r1", "api-case-040-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-041-gpt-6-luna-W-r1", "api-case-041-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-042-gpt-6-luna-W-r1", "api-case-042-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 2, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 1.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-043-gpt-6-luna-W-r1", "api-case-043-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-044-gpt-6-luna-W-r1", "api-case-044-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 2, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 1.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-045-gpt-6-luna-W-r1", "api-case-045-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-046-gpt-6-luna-W-r1", "api-case-046-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-047-gpt-6-luna-W-r1", "api-case-047-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-048-gpt-6-luna-W-r1", "api-case-048-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-049-gpt-6-luna-W-r1", "api-case-049-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-050-gpt-6-luna-W-r1", "api-case-050-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-051-gpt-6-luna-W-r1", "api-case-051-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-052-gpt-6-luna-W-r1", "api-case-052-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-053-gpt-6-luna-W-r1", "api-case-053-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-054-gpt-6-luna-W-r1", "api-case-054-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-055-gpt-6-luna-W-r1", "api-case-055-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-056-gpt-6-luna-W-r1", "api-case-056-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-057-gpt-6-luna-W-r1", "api-case-057-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-058-gpt-6-luna-W-r1", "api-case-058-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 2, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 1.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-059-gpt-6-luna-W-r1", "api-case-059-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-060-gpt-6-luna-W-r1", "api-case-060-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-061-gpt-6-luna-W-r1", "api-case-061-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "wrong_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-062-gpt-6-luna-W-r1", "api-case-062-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-063-gpt-6-luna-W-r1", "api-case-063-gpt-6-luna-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-064-gpt-6-luna-W-r1", "api-case-064-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-065-gpt-6-luna-W-r1", "api-case-065-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-066-gpt-6-luna-W-r1", "api-case-066-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-067-gpt-6-luna-W-r1", "api-case-067-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-068-gpt-6-luna-W-r1", "api-case-068-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-069-gpt-6-luna-W-r1", "api-case-069-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-070-gpt-6-luna-W-r1", "api-case-070-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-071-gpt-6-luna-W-r1", "api-case-071-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-072-gpt-6-luna-W-r1", "api-case-072-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-073-gpt-6-luna-W-r1", "api-case-073-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-074-gpt-6-luna-W-r1", "api-case-074-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-075-gpt-6-luna-W-r1", "api-case-075-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-076-gpt-6-luna-W-r1", "api-case-076-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-077-gpt-6-luna-W-r1", "api-case-077-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-078-gpt-6-luna-W-r1", "api-case-078-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-079-gpt-6-luna-W-r1", "api-case-079-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-080-gpt-6-luna-W-r1", "api-case-080-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-081-gpt-6-luna-W-r1", "api-case-081-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-082-gpt-6-luna-W-r1", "api-case-082-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-083-gpt-6-luna-W-r1", "api-case-083-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-084-gpt-6-luna-W-r1", "api-case-084-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-085-gpt-6-luna-W-r1", "api-case-085-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-086-gpt-6-luna-W-r1", "api-case-086-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-087-gpt-6-luna-W-r1", "api-case-087-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-088-gpt-6-luna-W-r1", "api-case-088-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-089-gpt-6-luna-W-r1", "api-case-089-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-090-gpt-6-luna-W-r1", "api-case-090-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-091-gpt-6-luna-W-r1", "api-case-091-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-092-gpt-6-luna-W-r1", "api-case-092-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-093-gpt-6-luna-W-r1", "api-case-093-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-094-gpt-6-luna-W-r1", "api-case-094-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-095-gpt-6-luna-W-r1", "api-case-095-gpt-6-luna-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6-luna", "arm": "W", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-096-gpt-6-luna-W-r1", "api-case-096-gpt-6-luna-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-001-gpt-6-1-sol-E-r1", "api-case-001-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-002-gpt-6-1-sol-E-r1", "api-case-002-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-003-gpt-6-1-sol-E-r1", "api-case-003-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-004-gpt-6-1-sol-E-r1", "api-case-004-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-005-gpt-6-1-sol-E-r1", "api-case-005-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-006-gpt-6-1-sol-E-r1", "api-case-006-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-007-gpt-6-1-sol-E-r1", "api-case-007-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-008-gpt-6-1-sol-E-r1", "api-case-008-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-009-gpt-6-1-sol-E-r1", "api-case-009-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-010-gpt-6-1-sol-E-r1", "api-case-010-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-011-gpt-6-1-sol-E-r1", "api-case-011-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-012-gpt-6-1-sol-E-r1", "api-case-012-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-013-gpt-6-1-sol-E-r1", "api-case-013-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-014-gpt-6-1-sol-E-r1", "api-case-014-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-015-gpt-6-1-sol-E-r1", "api-case-015-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-016-gpt-6-1-sol-E-r1", "api-case-016-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-017-gpt-6-1-sol-E-r1", "api-case-017-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-018-gpt-6-1-sol-E-r1", "api-case-018-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-019-gpt-6-1-sol-E-r1", "api-case-019-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-020-gpt-6-1-sol-E-r1", "api-case-020-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-021-gpt-6-1-sol-E-r1", "api-case-021-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-022-gpt-6-1-sol-E-r1", "api-case-022-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-023-gpt-6-1-sol-E-r1", "api-case-023-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-024-gpt-6-1-sol-E-r1", "api-case-024-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-025-gpt-6-1-sol-E-r1", "api-case-025-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-026-gpt-6-1-sol-E-r1", "api-case-026-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-027-gpt-6-1-sol-E-r1", "api-case-027-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-028-gpt-6-1-sol-E-r1", "api-case-028-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-029-gpt-6-1-sol-E-r1", "api-case-029-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-030-gpt-6-1-sol-E-r1", "api-case-030-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-031-gpt-6-1-sol-E-r1", "api-case-031-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-032-gpt-6-1-sol-E-r1", "api-case-032-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-033-gpt-6-1-sol-E-r1", "api-case-033-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-034-gpt-6-1-sol-E-r1", "api-case-034-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-035-gpt-6-1-sol-E-r1", "api-case-035-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-036-gpt-6-1-sol-E-r1", "api-case-036-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-037-gpt-6-1-sol-E-r1", "api-case-037-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-038-gpt-6-1-sol-E-r1", "api-case-038-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-039-gpt-6-1-sol-E-r1", "api-case-039-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-040-gpt-6-1-sol-E-r1", "api-case-040-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-041-gpt-6-1-sol-E-r1", "api-case-041-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-042-gpt-6-1-sol-E-r1", "api-case-042-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-043-gpt-6-1-sol-E-r1", "api-case-043-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-044-gpt-6-1-sol-E-r1", "api-case-044-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-045-gpt-6-1-sol-E-r1", "api-case-045-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-046-gpt-6-1-sol-E-r1", "api-case-046-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-047-gpt-6-1-sol-E-r1", "api-case-047-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-048-gpt-6-1-sol-E-r1", "api-case-048-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-049-gpt-6-1-sol-E-r1", "api-case-049-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-050-gpt-6-1-sol-E-r1", "api-case-050-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-051-gpt-6-1-sol-E-r1", "api-case-051-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-052-gpt-6-1-sol-E-r1", "api-case-052-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-053-gpt-6-1-sol-E-r1", "api-case-053-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-054-gpt-6-1-sol-E-r1", "api-case-054-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-055-gpt-6-1-sol-E-r1", "api-case-055-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-056-gpt-6-1-sol-E-r1", "api-case-056-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-057-gpt-6-1-sol-E-r1", "api-case-057-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-058-gpt-6-1-sol-E-r1", "api-case-058-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-059-gpt-6-1-sol-E-r1", "api-case-059-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-060-gpt-6-1-sol-E-r1", "api-case-060-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-061-gpt-6-1-sol-E-r1", "api-case-061-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-062-gpt-6-1-sol-E-r1", "api-case-062-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-063-gpt-6-1-sol-E-r1", "api-case-063-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-064-gpt-6-1-sol-E-r1", "api-case-064-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-065-gpt-6-1-sol-E-r1", "api-case-065-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-066-gpt-6-1-sol-E-r1", "api-case-066-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-067-gpt-6-1-sol-E-r1", "api-case-067-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-068-gpt-6-1-sol-E-r1", "api-case-068-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-069-gpt-6-1-sol-E-r1", "api-case-069-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-070-gpt-6-1-sol-E-r1", "api-case-070-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-071-gpt-6-1-sol-E-r1", "api-case-071-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-072-gpt-6-1-sol-E-r1", "api-case-072-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-073-gpt-6-1-sol-E-r1", "api-case-073-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-074-gpt-6-1-sol-E-r1", "api-case-074-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-075-gpt-6-1-sol-E-r1", "api-case-075-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-076-gpt-6-1-sol-E-r1", "api-case-076-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-077-gpt-6-1-sol-E-r1", "api-case-077-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-078-gpt-6-1-sol-E-r1", "api-case-078-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-079-gpt-6-1-sol-E-r1", "api-case-079-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-080-gpt-6-1-sol-E-r1", "api-case-080-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-081-gpt-6-1-sol-E-r1", "api-case-081-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-082-gpt-6-1-sol-E-r1", "api-case-082-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-083-gpt-6-1-sol-E-r1", "api-case-083-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-084-gpt-6-1-sol-E-r1", "api-case-084-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-085-gpt-6-1-sol-E-r1", "api-case-085-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-086-gpt-6-1-sol-E-r1", "api-case-086-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-087-gpt-6-1-sol-E-r1", "api-case-087-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-088-gpt-6-1-sol-E-r1", "api-case-088-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-089-gpt-6-1-sol-E-r1", "api-case-089-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-090-gpt-6-1-sol-E-r1", "api-case-090-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-091-gpt-6-1-sol-E-r1", "api-case-091-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-092-gpt-6-1-sol-E-r1", "api-case-092-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-093-gpt-6-1-sol-E-r1", "api-case-093-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-094-gpt-6-1-sol-E-r1", "api-case-094-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-095-gpt-6-1-sol-E-r1", "api-case-095-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "E", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-096-gpt-6-1-sol-E-r1", "api-case-096-gpt-6-1-sol-E-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-001-gpt-6-1-sol-M-r1", "api-case-001-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-002-gpt-6-1-sol-M-r1", "api-case-002-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-003-gpt-6-1-sol-M-r1", "api-case-003-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-004-gpt-6-1-sol-M-r1", "api-case-004-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-005-gpt-6-1-sol-M-r1", "api-case-005-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-006-gpt-6-1-sol-M-r1", "api-case-006-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-007-gpt-6-1-sol-M-r1", "api-case-007-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-008-gpt-6-1-sol-M-r1", "api-case-008-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-009-gpt-6-1-sol-M-r1", "api-case-009-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-010-gpt-6-1-sol-M-r1", "api-case-010-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-011-gpt-6-1-sol-M-r1", "api-case-011-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-012-gpt-6-1-sol-M-r1", "api-case-012-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-013-gpt-6-1-sol-M-r1", "api-case-013-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-014-gpt-6-1-sol-M-r1", "api-case-014-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-015-gpt-6-1-sol-M-r1", "api-case-015-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-016-gpt-6-1-sol-M-r1", "api-case-016-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-017-gpt-6-1-sol-M-r1", "api-case-017-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-018-gpt-6-1-sol-M-r1", "api-case-018-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-019-gpt-6-1-sol-M-r1", "api-case-019-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-020-gpt-6-1-sol-M-r1", "api-case-020-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-021-gpt-6-1-sol-M-r1", "api-case-021-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-022-gpt-6-1-sol-M-r1", "api-case-022-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-023-gpt-6-1-sol-M-r1", "api-case-023-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-024-gpt-6-1-sol-M-r1", "api-case-024-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-025-gpt-6-1-sol-M-r1", "api-case-025-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-026-gpt-6-1-sol-M-r1", "api-case-026-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-027-gpt-6-1-sol-M-r1", "api-case-027-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-028-gpt-6-1-sol-M-r1", "api-case-028-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-029-gpt-6-1-sol-M-r1", "api-case-029-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-030-gpt-6-1-sol-M-r1", "api-case-030-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-031-gpt-6-1-sol-M-r1", "api-case-031-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-032-gpt-6-1-sol-M-r1", "api-case-032-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-033-gpt-6-1-sol-M-r1", "api-case-033-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-034-gpt-6-1-sol-M-r1", "api-case-034-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-035-gpt-6-1-sol-M-r1", "api-case-035-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-036-gpt-6-1-sol-M-r1", "api-case-036-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-037-gpt-6-1-sol-M-r1", "api-case-037-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-038-gpt-6-1-sol-M-r1", "api-case-038-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-039-gpt-6-1-sol-M-r1", "api-case-039-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-040-gpt-6-1-sol-M-r1", "api-case-040-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-041-gpt-6-1-sol-M-r1", "api-case-041-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-042-gpt-6-1-sol-M-r1", "api-case-042-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-043-gpt-6-1-sol-M-r1", "api-case-043-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-044-gpt-6-1-sol-M-r1", "api-case-044-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-045-gpt-6-1-sol-M-r1", "api-case-045-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-046-gpt-6-1-sol-M-r1", "api-case-046-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-047-gpt-6-1-sol-M-r1", "api-case-047-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-048-gpt-6-1-sol-M-r1", "api-case-048-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-049-gpt-6-1-sol-M-r1", "api-case-049-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-050-gpt-6-1-sol-M-r1", "api-case-050-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-051-gpt-6-1-sol-M-r1", "api-case-051-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-052-gpt-6-1-sol-M-r1", "api-case-052-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-053-gpt-6-1-sol-M-r1", "api-case-053-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-054-gpt-6-1-sol-M-r1", "api-case-054-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-055-gpt-6-1-sol-M-r1", "api-case-055-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-056-gpt-6-1-sol-M-r1", "api-case-056-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-057-gpt-6-1-sol-M-r1", "api-case-057-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-058-gpt-6-1-sol-M-r1", "api-case-058-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-059-gpt-6-1-sol-M-r1", "api-case-059-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-060-gpt-6-1-sol-M-r1", "api-case-060-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-061-gpt-6-1-sol-M-r1", "api-case-061-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-062-gpt-6-1-sol-M-r1", "api-case-062-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-063-gpt-6-1-sol-M-r1", "api-case-063-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-064-gpt-6-1-sol-M-r1", "api-case-064-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-065-gpt-6-1-sol-M-r1", "api-case-065-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-066-gpt-6-1-sol-M-r1", "api-case-066-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-067-gpt-6-1-sol-M-r1", "api-case-067-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-068-gpt-6-1-sol-M-r1", "api-case-068-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-069-gpt-6-1-sol-M-r1", "api-case-069-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-070-gpt-6-1-sol-M-r1", "api-case-070-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-071-gpt-6-1-sol-M-r1", "api-case-071-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-072-gpt-6-1-sol-M-r1", "api-case-072-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-073-gpt-6-1-sol-M-r1", "api-case-073-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-074-gpt-6-1-sol-M-r1", "api-case-074-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-075-gpt-6-1-sol-M-r1", "api-case-075-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-076-gpt-6-1-sol-M-r1", "api-case-076-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-077-gpt-6-1-sol-M-r1", "api-case-077-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-078-gpt-6-1-sol-M-r1", "api-case-078-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-079-gpt-6-1-sol-M-r1", "api-case-079-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-080-gpt-6-1-sol-M-r1", "api-case-080-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-081-gpt-6-1-sol-M-r1", "api-case-081-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-082-gpt-6-1-sol-M-r1", "api-case-082-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-083-gpt-6-1-sol-M-r1", "api-case-083-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-084-gpt-6-1-sol-M-r1", "api-case-084-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-085-gpt-6-1-sol-M-r1", "api-case-085-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-086-gpt-6-1-sol-M-r1", "api-case-086-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-087-gpt-6-1-sol-M-r1", "api-case-087-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-088-gpt-6-1-sol-M-r1", "api-case-088-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-089-gpt-6-1-sol-M-r1", "api-case-089-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-090-gpt-6-1-sol-M-r1", "api-case-090-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-091-gpt-6-1-sol-M-r1", "api-case-091-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-092-gpt-6-1-sol-M-r1", "api-case-092-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-093-gpt-6-1-sol-M-r1", "api-case-093-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-094-gpt-6-1-sol-M-r1", "api-case-094-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-095-gpt-6-1-sol-M-r1", "api-case-095-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "M", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-096-gpt-6-1-sol-M-r1", "api-case-096-gpt-6-1-sol-M-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-001-gpt-6-1-sol-W-r1", "api-case-001-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-002-gpt-6-1-sol-W-r1", "api-case-002-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-003-gpt-6-1-sol-W-r1", "api-case-003-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-004-gpt-6-1-sol-W-r1", "api-case-004-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-005-gpt-6-1-sol-W-r1", "api-case-005-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-006-gpt-6-1-sol-W-r1", "api-case-006-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-007-gpt-6-1-sol-W-r1", "api-case-007-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-008-gpt-6-1-sol-W-r1", "api-case-008-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "request_ids": [ "api-case-009-gpt-6-1-sol-W-r1", "api-case-009-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-010-gpt-6-1-sol-W-r1", "api-case-010-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-011-gpt-6-1-sol-W-r1", "api-case-011-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-012-gpt-6-1-sol-W-r1", "api-case-012-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-013-gpt-6-1-sol-W-r1", "api-case-013-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-014-gpt-6-1-sol-W-r1", "api-case-014-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-015-gpt-6-1-sol-W-r1", "api-case-015-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-016-gpt-6-1-sol-W-r1", "api-case-016-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-017-gpt-6-1-sol-W-r1", "api-case-017-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-018-gpt-6-1-sol-W-r1", "api-case-018-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-019-gpt-6-1-sol-W-r1", "api-case-019-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-020-gpt-6-1-sol-W-r1", "api-case-020-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-021-gpt-6-1-sol-W-r1", "api-case-021-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-022-gpt-6-1-sol-W-r1", "api-case-022-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-023-gpt-6-1-sol-W-r1", "api-case-023-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-024-gpt-6-1-sol-W-r1", "api-case-024-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-025-gpt-6-1-sol-W-r1", "api-case-025-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-026-gpt-6-1-sol-W-r1", "api-case-026-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-027-gpt-6-1-sol-W-r1", "api-case-027-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-028-gpt-6-1-sol-W-r1", "api-case-028-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-029-gpt-6-1-sol-W-r1", "api-case-029-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-030-gpt-6-1-sol-W-r1", "api-case-030-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-031-gpt-6-1-sol-W-r1", "api-case-031-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-032-gpt-6-1-sol-W-r1", "api-case-032-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-033-gpt-6-1-sol-W-r1", "api-case-033-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-034-gpt-6-1-sol-W-r1", "api-case-034-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-035-gpt-6-1-sol-W-r1", "api-case-035-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "request_ids": [ "api-case-036-gpt-6-1-sol-W-r1", "api-case-036-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-037-gpt-6-1-sol-W-r1", "api-case-037-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-038-gpt-6-1-sol-W-r1", "api-case-038-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-039-gpt-6-1-sol-W-r1", "api-case-039-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-040-gpt-6-1-sol-W-r1", "api-case-040-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-041-gpt-6-1-sol-W-r1", "api-case-041-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-042-gpt-6-1-sol-W-r1", "api-case-042-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-043-gpt-6-1-sol-W-r1", "api-case-043-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 1, "at_least_one_wrong_assertion": true, "wrong_assertion_mean": 0.5, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-044-gpt-6-1-sol-W-r1", "api-case-044-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-045-gpt-6-1-sol-W-r1", "api-case-045-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-046-gpt-6-1-sol-W-r1", "api-case-046-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-047-gpt-6-1-sol-W-r1", "api-case-047-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-048-gpt-6-1-sol-W-r1", "api-case-048-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-049-gpt-6-1-sol-W-r1", "api-case-049-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-050-gpt-6-1-sol-W-r1", "api-case-050-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-051-gpt-6-1-sol-W-r1", "api-case-051-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-052-gpt-6-1-sol-W-r1", "api-case-052-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-053-gpt-6-1-sol-W-r1", "api-case-053-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-054-gpt-6-1-sol-W-r1", "api-case-054-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-055-gpt-6-1-sol-W-r1", "api-case-055-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-056-gpt-6-1-sol-W-r1", "api-case-056-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-057-gpt-6-1-sol-W-r1", "api-case-057-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-058-gpt-6-1-sol-W-r1", "api-case-058-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-059-gpt-6-1-sol-W-r1", "api-case-059-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-060-gpt-6-1-sol-W-r1", "api-case-060-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-061-gpt-6-1-sol-W-r1", "api-case-061-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_absent_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-062-gpt-6-1-sol-W-r1", "api-case-062-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-063-gpt-6-1-sol-W-r1", "api-case-063-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_absent_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "request_ids": [ "api-case-064-gpt-6-1-sol-W-r1", "api-case-064-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-065-gpt-6-1-sol-W-r1", "api-case-065-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-066-gpt-6-1-sol-W-r1", "api-case-066-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-067-gpt-6-1-sol-W-r1", "api-case-067-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-068-gpt-6-1-sol-W-r1", "api-case-068-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-069-gpt-6-1-sol-W-r1", "api-case-069-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-070-gpt-6-1-sol-W-r1", "api-case-070-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-071-gpt-6-1-sol-W-r1", "api-case-071-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-072-gpt-6-1-sol-W-r1", "api-case-072-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-073-gpt-6-1-sol-W-r1", "api-case-073-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-074-gpt-6-1-sol-W-r1", "api-case-074-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-075-gpt-6-1-sol-W-r1", "api-case-075-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-076-gpt-6-1-sol-W-r1", "api-case-076-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-077-gpt-6-1-sol-W-r1", "api-case-077-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-078-gpt-6-1-sol-W-r1", "api-case-078-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-079-gpt-6-1-sol-W-r1", "api-case-079-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-080-gpt-6-1-sol-W-r1", "api-case-080-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-081-gpt-6-1-sol-W-r1", "api-case-081-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-082-gpt-6-1-sol-W-r1", "api-case-082-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-083-gpt-6-1-sol-W-r1", "api-case-083-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "request_ids": [ "api-case-084-gpt-6-1-sol-W-r1", "api-case-084-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "unresolved" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-085-gpt-6-1-sol-W-r1", "api-case-085-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-086-gpt-6-1-sol-W-r1", "api-case-086-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-087-gpt-6-1-sol-W-r1", "api-case-087-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-088-gpt-6-1-sol-W-r1", "api-case-088-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-089-gpt-6-1-sol-W-r1", "api-case-089-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-090-gpt-6-1-sol-W-r1", "api-case-090-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "unresolved", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-091-gpt-6-1-sol-W-r1", "api-case-091-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "request_ids": [ "api-case-092-gpt-6-1-sol-W-r1", "api-case-092-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-093-gpt-6-1-sol-W-r1", "api-case-093-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-094-gpt-6-1-sol-W-r1", "api-case-094-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-095-gpt-6-1-sol-W-r1", "api-case-095-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true }, { "model": "gpt-6.1-sol", "arm": "W", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "request_ids": [ "api-case-096-gpt-6-1-sol-W-r1", "api-case-096-gpt-6-1-sol-W-r2" ], "membership_outcomes": [ "correct_present_claim", "correct_present_claim" ], "operational_outcomes": [ "complete", "complete" ], "exactly_two_complete_repetitions": true, "wrong_assertions_in_two": 0, "at_least_one_wrong_assertion": false, "wrong_assertion_mean": 0.0, "full_text_reviewed_both": true } ], "company_metrics_by_model_arm_cohort": [ { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "selected_companies": 32, "companies_with_exactly_two_complete_repetitions": 32, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "selected_companies": 64, "companies_with_exactly_two_complete_repetitions": 64, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "selected_companies": 32, "companies_with_exactly_two_complete_repetitions": 32, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "selected_companies": 64, "companies_with_exactly_two_complete_repetitions": 64, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "selected_companies": 32, "companies_with_exactly_two_complete_repetitions": 32, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "selected_companies": 64, "companies_with_exactly_two_complete_repetitions": 64, "companies_with_at_least_one_wrong_assertion": 23, "companies_wrong_in_both_repetitions": 5, "mean_of_company_wrong_assertion_means": 0.21875, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "selected_companies": 32, "companies_with_exactly_two_complete_repetitions": 32, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "selected_companies": 64, "companies_with_exactly_two_complete_repetitions": 64, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "selected_companies": 32, "companies_with_exactly_two_complete_repetitions": 32, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "selected_companies": 64, "companies_with_exactly_two_complete_repetitions": 64, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "selected_companies": 32, "companies_with_exactly_two_complete_repetitions": 32, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "selected_companies": 64, "companies_with_exactly_two_complete_repetitions": 64, "companies_with_at_least_one_wrong_assertion": 1, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0078125, "companies_missing_complete_pair": 0 } ], "company_metrics_by_model_arm_cohort_event_stratum": [ { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "selected_companies": 8, "companies_with_exactly_two_complete_repetitions": 8, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "selected_companies": 20, "companies_with_exactly_two_complete_repetitions": 20, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "selected_companies": 4, "companies_with_exactly_two_complete_repetitions": 4, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "selected_companies": 9, "companies_with_exactly_two_complete_repetitions": 9, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "selected_companies": 27, "companies_with_exactly_two_complete_repetitions": 27, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "E", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "selected_companies": 28, "companies_with_exactly_two_complete_repetitions": 28, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "selected_companies": 8, "companies_with_exactly_two_complete_repetitions": 8, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "selected_companies": 20, "companies_with_exactly_two_complete_repetitions": 20, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "selected_companies": 4, "companies_with_exactly_two_complete_repetitions": 4, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "selected_companies": 9, "companies_with_exactly_two_complete_repetitions": 9, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "selected_companies": 27, "companies_with_exactly_two_complete_repetitions": 27, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "M", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "selected_companies": 28, "companies_with_exactly_two_complete_repetitions": 28, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "selected_companies": 8, "companies_with_exactly_two_complete_repetitions": 8, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "selected_companies": 20, "companies_with_exactly_two_complete_repetitions": 20, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "selected_companies": 4, "companies_with_exactly_two_complete_repetitions": 4, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "selected_companies": 9, "companies_with_exactly_two_complete_repetitions": 9, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "selected_companies": 27, "companies_with_exactly_two_complete_repetitions": 27, "companies_with_at_least_one_wrong_assertion": 8, "companies_wrong_in_both_repetitions": 1, "mean_of_company_wrong_assertion_means": 0.16666666666666666, "companies_missing_complete_pair": 0 }, { "model": "gpt-6-luna", "arm": "W", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "selected_companies": 28, "companies_with_exactly_two_complete_repetitions": 28, "companies_with_at_least_one_wrong_assertion": 15, "companies_wrong_in_both_repetitions": 4, "mean_of_company_wrong_assertion_means": 0.3392857142857143, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "selected_companies": 8, "companies_with_exactly_two_complete_repetitions": 8, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "selected_companies": 20, "companies_with_exactly_two_complete_repetitions": 20, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "selected_companies": 4, "companies_with_exactly_two_complete_repetitions": 4, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "selected_companies": 9, "companies_with_exactly_two_complete_repetitions": 9, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "selected_companies": 27, "companies_with_exactly_two_complete_repetitions": 27, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "E", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "selected_companies": 28, "companies_with_exactly_two_complete_repetitions": 28, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "selected_companies": 8, "companies_with_exactly_two_complete_repetitions": 8, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "selected_companies": 20, "companies_with_exactly_two_complete_repetitions": 20, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "selected_companies": 4, "companies_with_exactly_two_complete_repetitions": 4, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "selected_companies": 9, "companies_with_exactly_two_complete_repetitions": 9, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "selected_companies": 27, "companies_with_exactly_two_complete_repetitions": 27, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "M", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "selected_companies": 28, "companies_with_exactly_two_complete_repetitions": 28, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-IRAQ2", "selected_companies": 8, "companies_with_exactly_two_complete_repetitions": 8, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNT", "selected_companies": 20, "companies_with_exactly_two_complete_repetitions": 20, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "current_control", "event": null, "stratum": "listed-control-SDNTK", "selected_companies": 4, "companies_with_exactly_two_complete_repetitions": 4, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "event": "2026-05-28", "stratum": "removed-20260528", "selected_companies": 9, "companies_with_exactly_two_complete_repetitions": 9, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "event": "2026-07-27", "stratum": "removed-20260727", "selected_companies": 27, "companies_with_exactly_two_complete_repetitions": 27, "companies_with_at_least_one_wrong_assertion": 0, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.0, "companies_missing_complete_pair": 0 }, { "model": "gpt-6.1-sol", "arm": "W", "cohort": "removed", "event": "2026-10-05", "stratum": "removed-20261005", "selected_companies": 28, "companies_with_exactly_two_complete_repetitions": 28, "companies_with_at_least_one_wrong_assertion": 1, "companies_wrong_in_both_repetitions": 0, "mean_of_company_wrong_assertion_means": 0.017857142857142856, "companies_missing_complete_pair": 0 } ], "paired_comparisons": { "company_pairs": [ { "model": "gpt-6-luna", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 1.0, "difference_right_minus_left": 1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 1.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 1.0, "difference_right_minus_left": 1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 1.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 1.0, "difference_right_minus_left": 1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 1.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 1.0, "difference_right_minus_left": 1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 1.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 1.0, "difference_right_minus_left": 1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 1.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -1.0 }, { "model": "gpt-6-luna", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "wrong_present_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "wrong_present_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6-luna", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-001", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-002", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-003", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-004", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-005", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-006", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-007", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-008", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-009", "cohort": "removed", "stratum": "removed-20260528", "event": "2026-05-28", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-010", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-011", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-012", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-013", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-014", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-015", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-016", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-017", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-018", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-019", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-020", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-021", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-022", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-023", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-024", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-025", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-026", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-027", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-028", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-029", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-030", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-031", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-032", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-033", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-034", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-035", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-036", "cohort": "removed", "stratum": "removed-20260727", "event": "2026-07-27", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-037", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-038", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-039", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-040", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-041", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-042", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": true, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.5, "difference_right_minus_left": 0.5 }, { "model": "gpt-6.1-sol", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "wrong_present_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": true, "right_any_wrong": false, "left_wrong_assertion_mean": 0.5, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": -0.5 }, { "model": "gpt-6.1-sol", "case_id": "api-case-043", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-044", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-045", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-046", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-047", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-048", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-049", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-050", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-051", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-052", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-053", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-054", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-055", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-056", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-057", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-058", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-059", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-060", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_absent_claim" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-061", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-062", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_absent_claim", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-063", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-064", "cohort": "removed", "stratum": "removed-20261005", "event": "2026-10-05", "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_absent_claim", "correct_absent_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-065", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-066", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-067", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-068", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-069", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-070", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-071", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-072", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-073", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-074", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-075", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-076", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-077", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-078", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-079", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-080", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-081", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-082", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-083", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "unresolved" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-084", "cohort": "current_control", "stratum": "listed-control-SDNT", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-085", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-086", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-087", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-088", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-089", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "unresolved", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-090", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-091", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-092", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-093", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-094", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-095", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_W", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "W_vs_E", "comparison_role": "prespecified", "both_arms_two_complete": true, "left_outcomes": [ "correct_present_claim", "correct_present_claim" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "case_id": "api-case-096", "cohort": "current_control", "stratum": "listed-control-SDNTK", "event": null, "comparison": "M_vs_E", "comparison_role": "secondary", "both_arms_two_complete": true, "left_outcomes": [ "unresolved", "unresolved" ], "right_outcomes": [ "correct_present_claim", "correct_present_claim" ], "left_any_wrong": false, "right_any_wrong": false, "left_wrong_assertion_mean": 0.0, "right_wrong_assertion_mean": 0.0, "difference_right_minus_left": 0.0 } ], "by_model_comparison_cohort_stratum": [ { "model": "gpt-6-luna", "comparison": "M_vs_E", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "selected_companies": 8, "complete_paired_companies": 8, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 8 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNT", "selected_companies": 20, "complete_paired_companies": 20, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 20 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNTK", "selected_companies": 4, "complete_paired_companies": 4, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 4 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_E", "cohort": "removed", "stratum": "removed-20260528", "selected_companies": 9, "complete_paired_companies": 9, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 9 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_E", "cohort": "removed", "stratum": "removed-20260727", "selected_companies": 27, "complete_paired_companies": 27, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 27 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_E", "cohort": "removed", "stratum": "removed-20261005", "selected_companies": 28, "complete_paired_companies": 28, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 28 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_W", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "selected_companies": 8, "complete_paired_companies": 8, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 8 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_W", "cohort": "current_control", "stratum": "listed-control-SDNT", "selected_companies": 20, "complete_paired_companies": 20, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 20 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_W", "cohort": "current_control", "stratum": "listed-control-SDNTK", "selected_companies": 4, "complete_paired_companies": 4, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 4 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_W", "cohort": "removed", "stratum": "removed-20260528", "selected_companies": 9, "complete_paired_companies": 9, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 9 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "M_vs_W", "cohort": "removed", "stratum": "removed-20260727", "selected_companies": 27, "complete_paired_companies": 27, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 19, "none_wrong_to_any_wrong": 8 }, "mean_company_difference_right_minus_left": 0.16666666666666666 }, { "model": "gpt-6-luna", "comparison": "M_vs_W", "cohort": "removed", "stratum": "removed-20261005", "selected_companies": 28, "complete_paired_companies": 28, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_any_wrong": 15, "none_wrong_to_none_wrong": 13 }, "mean_company_difference_right_minus_left": 0.3392857142857143 }, { "model": "gpt-6-luna", "comparison": "W_vs_E", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "selected_companies": 8, "complete_paired_companies": 8, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 8 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "W_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNT", "selected_companies": 20, "complete_paired_companies": 20, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 20 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "W_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNTK", "selected_companies": 4, "complete_paired_companies": 4, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 4 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "W_vs_E", "cohort": "removed", "stratum": "removed-20260528", "selected_companies": 9, "complete_paired_companies": 9, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 9 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6-luna", "comparison": "W_vs_E", "cohort": "removed", "stratum": "removed-20260727", "selected_companies": 27, "complete_paired_companies": 27, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 19, "any_wrong_to_none_wrong": 8 }, "mean_company_difference_right_minus_left": -0.16666666666666666 }, { "model": "gpt-6-luna", "comparison": "W_vs_E", "cohort": "removed", "stratum": "removed-20261005", "selected_companies": 28, "complete_paired_companies": 28, "incomplete_paired_companies": 0, "any_wrong_transitions": { "any_wrong_to_none_wrong": 15, "none_wrong_to_none_wrong": 13 }, "mean_company_difference_right_minus_left": -0.3392857142857143 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_E", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "selected_companies": 8, "complete_paired_companies": 8, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 8 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNT", "selected_companies": 20, "complete_paired_companies": 20, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 20 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNTK", "selected_companies": 4, "complete_paired_companies": 4, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 4 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_E", "cohort": "removed", "stratum": "removed-20260528", "selected_companies": 9, "complete_paired_companies": 9, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 9 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_E", "cohort": "removed", "stratum": "removed-20260727", "selected_companies": 27, "complete_paired_companies": 27, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 27 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_E", "cohort": "removed", "stratum": "removed-20261005", "selected_companies": 28, "complete_paired_companies": 28, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 28 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_W", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "selected_companies": 8, "complete_paired_companies": 8, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 8 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_W", "cohort": "current_control", "stratum": "listed-control-SDNT", "selected_companies": 20, "complete_paired_companies": 20, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 20 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_W", "cohort": "current_control", "stratum": "listed-control-SDNTK", "selected_companies": 4, "complete_paired_companies": 4, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 4 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_W", "cohort": "removed", "stratum": "removed-20260528", "selected_companies": 9, "complete_paired_companies": 9, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 9 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_W", "cohort": "removed", "stratum": "removed-20260727", "selected_companies": 27, "complete_paired_companies": 27, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 27 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "M_vs_W", "cohort": "removed", "stratum": "removed-20261005", "selected_companies": 28, "complete_paired_companies": 28, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 27, "none_wrong_to_any_wrong": 1 }, "mean_company_difference_right_minus_left": 0.017857142857142856 }, { "model": "gpt-6.1-sol", "comparison": "W_vs_E", "cohort": "current_control", "stratum": "listed-control-IRAQ2", "selected_companies": 8, "complete_paired_companies": 8, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 8 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "W_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNT", "selected_companies": 20, "complete_paired_companies": 20, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 20 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "W_vs_E", "cohort": "current_control", "stratum": "listed-control-SDNTK", "selected_companies": 4, "complete_paired_companies": 4, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 4 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "W_vs_E", "cohort": "removed", "stratum": "removed-20260528", "selected_companies": 9, "complete_paired_companies": 9, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 9 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "W_vs_E", "cohort": "removed", "stratum": "removed-20260727", "selected_companies": 27, "complete_paired_companies": 27, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 27 }, "mean_company_difference_right_minus_left": 0.0 }, { "model": "gpt-6.1-sol", "comparison": "W_vs_E", "cohort": "removed", "stratum": "removed-20261005", "selected_companies": 28, "complete_paired_companies": 28, "incomplete_paired_companies": 0, "any_wrong_transitions": { "none_wrong_to_none_wrong": 27, "any_wrong_to_none_wrong": 1 }, "mean_company_difference_right_minus_left": -0.017857142857142856 } ], "interpretation": "None wrong may include unresolved answers. It does not mean both answers were correct. Comparisons require two complete repetitions in each arm." }, "limits": [ "Selected companies and repeated responses are distinct units. No population accuracy claim, confidence interval, or independent-observation p value is calculated.", "This is a purposive event panel, with repeated requests nested in companies and only three selected removal events.", "No-tool means the tested API request had no external tools. It does not describe the consumer ChatGPT product or all AI.", "Supplied-evidence results measure comprehension of supplied evidence, not autonomous retrieval, hosted product performance, or sanctions clearance.", "Source URLs observed in traces are not proof that the cited content supports the assertion. Citation problems do not automatically change membership correctness.", "web_call_count and web_calls_observed count returned web_search_call items, including uncompleted searching attempts. completed_web_action_count records only completed items. These are distinct from provider-reported billable search requests.", "Automatic categories follow the structured verdict; explanation conflicts and final corrections require documented full-answer review.", "Missing response files in a running snapshot are pending or ambiguous, never treated as wrong answers.", "No attempt to decode or publish provider reasoning ciphertext is made. Only visible messages and tool-source metadata are extracted." ], "publication_checks": { "approved_for_publication": true, "approved_at_utc": "2026-10-08T17:50:27.582755+00:00", "approval_scope": "Internal research publication check; not external peer review or human expert certification", "scheduled_requests": 1152, "complete_and_full_text_reviewed_requests": 1152, "operational_failures": 0, "consequential_wrong_claims_corroborated": 29, "official_source_unchanged_through_end_check": true, "official_source_end_check_utc": "2026-10-08T17:48:43.657357+00:00", "official_source_sha256": "259336aa866dda3fc0cd3cddb933f4b5d48025d47035bcdece9b4ec90bf7ed92", "production_read_only_audit": { "scope": "Read-only production public-source reconciliation. No hosted screening request, fuzzy-match benchmark, or AI-plus-SanctionsKit integration tested.", "observed_at_utc": "2026-10-08T17:48:44.634Z", "source_retrieved_at": "2026-10-08T17:11:22.484Z", "product_published_at": "2026-10-08T00:00:00.000Z", "product_last_checked_at": "2026-10-08T17:11:22.484Z", "product_source_sha256": "259336aa866dda3fc0cd3cddb933f4b5d48025d47035bcdece9b4ec90bf7ed92", "official_source_sha256": "259336aa866dda3fc0cd3cddb933f4b5d48025d47035bcdece9b4ec90bf7ed92", "official_publication": "10/08/2026", "official_records": 19416, "product_records": 19416, "source_hashes_equal": true, "uid_inventory_equal": true, "missing_product_uids": [], "extra_product_uids": [], "selected_companies": 96, "selected_expected_membership_agreement": 96, "selected_all_query_agreement": 96, "selected_common_record_field_agreement": 96, "audit_script_sha256": "7a84f23baad9d1a7ed4f467fa08d02eabe6312f81bb357ce3dafc6060e9d98df", "record_field_comparison": "Official Entity maps to product organization; names, identifiers, programs, and country fields compared after the documented common projection.", "cases": [ { "case_id": "api-case-001", "company_name": "DISTRIBUIDORA DE ELEMENTOS PARA LA CONSTRUCCION S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "DISTRIBUIDORA DE ELEMENTOS PARA LA CONSTRUCCION S.A.", "normalized": "distribuidoradeelementosparalaconstruccionsa", "matched_uids": [] }, { "query": "D'ELCON S.A.", "normalized": "delconsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800117780-2", "source_qualifier": "Colombia" }, "key": [ "nit", "8001177802" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "DISTRIBUIDORA DE ELEMENTOS PARA LA CONSTRUCCION S.A.", "normalized": "distribuidoradeelementosparalaconstruccionsa", "matched_uids": [] }, { "query": "D'ELCON S.A.", "normalized": "delconsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800117780-2", "source_qualifier": "Colombia" }, "key": [ "nit", "8001177802" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-002", "company_name": "COMERCIALIZADORA DE CARNES LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "COMERCIALIZADORA DE CARNES LTDA.", "normalized": "comercializadoradecarnesltda", "matched_uids": [] }, { "query": "COMECARNES LTDA.", "normalized": "comecarnesltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800076369-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8000763690" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "COMERCIALIZADORA DE CARNES LTDA.", "normalized": "comercializadoradecarnesltda", "matched_uids": [] }, { "query": "COMECARNES LTDA.", "normalized": "comecarnesltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800076369-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8000763690" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-003", "company_name": "COLOMBIANA DE CERDOS LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "COLOMBIANA DE CERDOS LTDA.", "normalized": "colombianadecerdosltda", "matched_uids": [] }, { "query": "COLCERDOS LTDA.", "normalized": "colcerdosltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800018928-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8000189280" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "COLOMBIANA DE CERDOS LTDA.", "normalized": "colombianadecerdosltda", "matched_uids": [] }, { "query": "COLCERDOS LTDA.", "normalized": "colcerdosltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800018928-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8000189280" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-004", "company_name": "MATADERO METROPOLITANO LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "MATADERO METROPOLITANO LTDA.", "normalized": "mataderometropolitanoltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "891412986-8", "source_qualifier": "Colombia" }, "key": [ "nit", "8914129868" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "MATADERO METROPOLITANO LTDA.", "normalized": "mataderometropolitanoltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "891412986-8", "source_qualifier": "Colombia" }, "key": [ "nit", "8914129868" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-005", "company_name": "SERVIAUTOS UNO A 1A LIMITADA", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "SERVIAUTOS UNO A 1A LIMITADA", "normalized": "serviautosunoa1alimitada", "matched_uids": [] }, { "query": "DIAGNOSTICENTRO LA GARANTIA", "normalized": "diagnosticentrolagarantia", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800032413-8", "source_qualifier": "Colombia" }, "key": [ "nit", "8000324138" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "SERVIAUTOS UNO A 1A LIMITADA", "normalized": "serviautosunoa1alimitada", "matched_uids": [] }, { "query": "DIAGNOSTICENTRO LA GARANTIA", "normalized": "diagnosticentrolagarantia", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800032413-8", "source_qualifier": "Colombia" }, "key": [ "nit", "8000324138" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-006", "company_name": "CAVIEDES DILEO Y CIA. S.C.S.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "CAVIEDES DILEO Y CIA. S.C.S.", "normalized": "caviedesdileoyciascs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800113437-2", "source_qualifier": "Colombia" }, "key": [ "nit", "8001134372" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "CAVIEDES DILEO Y CIA. S.C.S.", "normalized": "caviedesdileoyciascs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800113437-2", "source_qualifier": "Colombia" }, "key": [ "nit", "8001134372" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-007", "company_name": "CONSTRUCTORA AMERICA S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "CONSTRUCTORA AMERICA S.A.", "normalized": "constructoraamericasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830125002-3", "source_qualifier": "Colombia" }, "key": [ "nit", "8301250023" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "CONSTRUCTORA AMERICA S.A.", "normalized": "constructoraamericasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830125002-3", "source_qualifier": "Colombia" }, "key": [ "nit", "8301250023" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-008", "company_name": "INVERSIONES Y REPRESENTACIONES S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "INVERSIONES Y REPRESENTACIONES S.A.", "normalized": "inversionesyrepresentacionessa", "matched_uids": [] }, { "query": "IRSA S.A.", "normalized": "irsasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "811040270-5", "source_qualifier": "Colombia" }, "key": [ "nit", "8110402705" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "INVERSIONES Y REPRESENTACIONES S.A.", "normalized": "inversionesyrepresentacionessa", "matched_uids": [] }, { "query": "IRSA S.A.", "normalized": "irsasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "811040270-5", "source_qualifier": "Colombia" }, "key": [ "nit", "8110402705" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-009", "company_name": "AGROPECUARIA PALMA DEL RIO S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "AGROPECUARIA PALMA DEL RIO S.A.", "normalized": "agropecuariapalmadelriosa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830061299-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8300612997" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "AGROPECUARIA PALMA DEL RIO S.A.", "normalized": "agropecuariapalmadelriosa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830061299-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8300612997" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-010", "company_name": "INVESTACAST PRECISION CASTINGS, LTD.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "INVESTACAST PRECISION CASTINGS, LTD.", "normalized": "investacastprecisioncastingsltd", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "INVESTACAST PRECISION CASTINGS, LTD.", "normalized": "investacastprecisioncastingsltd", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-011", "company_name": "U.I. INTERNATIONAL", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "U.I. INTERNATIONAL", "normalized": "uiinternational", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "U.I. INTERNATIONAL", "normalized": "uiinternational", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-012", "company_name": "TRADING & MARITIME INVESTMENTS", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "TRADING & MARITIME INVESTMENTS", "normalized": "tradingandmaritimeinvestments", "matched_uids": [] }, { "query": "TRADING AND MARITIME INVESTMENTS", "normalized": "tradingandmaritimeinvestments", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "TRADING & MARITIME INVESTMENTS", "normalized": "tradingandmaritimeinvestments", "matched_uids": [] }, { "query": "TRADING AND MARITIME INVESTMENTS", "normalized": "tradingandmaritimeinvestments", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-013", "company_name": "A.T.E. INTERNATIONAL LTD.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "A.T.E. INTERNATIONAL LTD.", "normalized": "ateinternationalltd", "matched_uids": [] }, { "query": "RWR INTERNATIONAL COMMODITIES", "normalized": "rwrinternationalcommodities", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "A.T.E. INTERNATIONAL LTD.", "normalized": "ateinternationalltd", "matched_uids": [] }, { "query": "RWR INTERNATIONAL COMMODITIES", "normalized": "rwrinternationalcommodities", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-014", "company_name": "ADVANCED ELECTRONICS DEVELOPMENT, LTD", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ADVANCED ELECTRONICS DEVELOPMENT, LTD", "normalized": "advancedelectronicsdevelopmentltd", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ADVANCED ELECTRONICS DEVELOPMENT, LTD", "normalized": "advancedelectronicsdevelopmentltd", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-015", "company_name": "FALCON SYSTEMS", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "FALCON SYSTEMS", "normalized": "falconsystems", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "FALCON SYSTEMS", "normalized": "falconsystems", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-016", "company_name": "IRAQI TRADE CENTER", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "IRAQI TRADE CENTER", "normalized": "iraqitradecenter", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "IRAQI TRADE CENTER", "normalized": "iraqitradecenter", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-017", "company_name": "A.W.A. ENGINEERING LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "A.W.A. ENGINEERING LIMITED", "normalized": "awaengineeringlimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "A.W.A. ENGINEERING LIMITED", "normalized": "awaengineeringlimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-018", "company_name": "INVERSIONES C.P.C.L. Y CIA. S. EN C.S.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "INVERSIONES C.P.C.L. Y CIA. S. EN C.S.", "normalized": "inversionescpclyciasencs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "900315175-8", "source_qualifier": "Colombia" }, "key": [ "nit", "9003151758" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "INVERSIONES C.P.C.L. Y CIA. S. EN C.S.", "normalized": "inversionescpclyciasencs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "900315175-8", "source_qualifier": "Colombia" }, "key": [ "nit", "9003151758" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-019", "company_name": "ARIZONA S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ARIZONA S.A.", "normalized": "arizonasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "836000489-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8360004890" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ARIZONA S.A.", "normalized": "arizonasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "836000489-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8360004890" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-020", "company_name": "ARCHI CENTRE I.C.E. LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ARCHI CENTRE I.C.E. LIMITED", "normalized": "archicentreicelimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ARCHI CENTRE I.C.E. LIMITED", "normalized": "archicentreicelimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-021", "company_name": "AGRICOLA GANADERA HENAO GONZALEZ Y CIA. S.C.S.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "AGRICOLA GANADERA HENAO GONZALEZ Y CIA. S.C.S.", "normalized": "agricolaganaderahenaogonzalezyciascs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800021615-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8000216151" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "AGRICOLA GANADERA HENAO GONZALEZ Y CIA. S.C.S.", "normalized": "agricolaganaderahenaogonzalezyciascs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800021615-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8000216151" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-022", "company_name": "ATLAS EQUIPMENT COMPANY LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ATLAS EQUIPMENT COMPANY LIMITED", "normalized": "atlasequipmentcompanylimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ATLAS EQUIPMENT COMPANY LIMITED", "normalized": "atlasequipmentcompanylimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-023", "company_name": "ADMINCHECK LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ADMINCHECK LIMITED", "normalized": "adminchecklimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ADMINCHECK LIMITED", "normalized": "adminchecklimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-024", "company_name": "KEENCLOUD LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "KEENCLOUD LIMITED", "normalized": "keencloudlimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "KEENCLOUD LIMITED", "normalized": "keencloudlimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-025", "company_name": "COMPANIA AGROINVERSORA HENAGRO LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "COMPANIA AGROINVERSORA HENAGRO LTDA.", "normalized": "companiaagroinversorahenagroltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800084326-8", "source_qualifier": "Colombia" }, "key": [ "nit", "8000843268" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "COMPANIA AGROINVERSORA HENAGRO LTDA.", "normalized": "companiaagroinversorahenagroltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800084326-8", "source_qualifier": "Colombia" }, "key": [ "nit", "8000843268" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-026", "company_name": "ORIENT SHIPPING LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ORIENT SHIPPING LIMITED", "normalized": "orientshippinglimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ORIENT SHIPPING LIMITED", "normalized": "orientshippinglimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-027", "company_name": "MEED INTERNATIONAL LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "MEED INTERNATIONAL LIMITED", "normalized": "meedinternationallimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "MEED INTERNATIONAL LIMITED", "normalized": "meedinternationallimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-028", "company_name": "JARACO S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "JARACO S.A.", "normalized": "jaracosa", "matched_uids": [] }, { "query": "SOKTAR", "normalized": "soktar", "matched_uids": [] }, { "query": "TRADACO S.A.", "normalized": "tradacosa", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "JARACO S.A.", "normalized": "jaracosa", "matched_uids": [] }, { "query": "SOKTAR", "normalized": "soktar", "matched_uids": [] }, { "query": "TRADACO S.A.", "normalized": "tradacosa", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-029", "company_name": "ORGANIZACION EMPRESARIAL A DE J HENAO M E HIJOS Y CIA. S.C.S.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ORGANIZACION EMPRESARIAL A DE J HENAO M E HIJOS Y CIA. S.C.S.", "normalized": "organizacionempresarialadejhenaomehijosyciascs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800157331-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8001573311" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ORGANIZACION EMPRESARIAL A DE J HENAO M E HIJOS Y CIA. S.C.S.", "normalized": "organizacionempresarialadejhenaomehijosyciascs", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800157331-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8001573311" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-030", "company_name": "T.E.G. LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "T.E.G. LIMITED", "normalized": "teglimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "T.E.G. LIMITED", "normalized": "teglimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-031", "company_name": "MAQUINARIA TECNICA Y TIERRAS LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "MAQUINARIA TECNICA Y TIERRAS LTDA.", "normalized": "maquinariatecnicaytierrasltda", "matched_uids": [] }, { "query": "M.T.T. LTDA.", "normalized": "mttltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800084233-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8000842331" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "MAQUINARIA TECNICA Y TIERRAS LTDA.", "normalized": "maquinariatecnicaytierrasltda", "matched_uids": [] }, { "query": "M.T.T. LTDA.", "normalized": "mttltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800084233-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8000842331" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-032", "company_name": "ATLAS AIR CONDITIONING COMPANY LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ATLAS AIR CONDITIONING COMPANY LIMITED", "normalized": "atlasairconditioningcompanylimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ATLAS AIR CONDITIONING COMPANY LIMITED", "normalized": "atlasairconditioningcompanylimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-033", "company_name": "LA TIENDA DE MINGO", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "LA TIENDA DE MINGO", "normalized": "latiendademingo", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "Matricula Mercantil No", "value": "16218702", "source_qualifier": "Medellin" }, "key": [ "matriculamercantil", "16218702" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "LA TIENDA DE MINGO", "normalized": "latiendademingo", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "Matricula Mercantil No", "value": "16218702", "source_qualifier": "Medellin" }, "key": [ "matriculamercantil", "16218702" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-034", "company_name": "ASSOCIATED ENGINEERS", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ASSOCIATED ENGINEERS", "normalized": "associatedengineers", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ASSOCIATED ENGINEERS", "normalized": "associatedengineers", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-035", "company_name": "T N K FABRICS LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "T N K FABRICS LIMITED", "normalized": "tnkfabricslimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "T N K FABRICS LIMITED", "normalized": "tnkfabricslimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-036", "company_name": "ENDSHIRE EXPORT MARKETING", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ENDSHIRE EXPORT MARKETING", "normalized": "endshireexportmarketing", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ENDSHIRE EXPORT MARKETING", "normalized": "endshireexportmarketing", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-037", "company_name": "GRUPO GUADALEST S.A. DE C.V.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "GRUPO GUADALEST S.A. DE C.V.", "normalized": "grupoguadalestsadecv", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "R.F.C.", "value": "GGU040603B20", "source_qualifier": "Mexico" }, "key": [ "rfc", "ggu040603b20" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "GRUPO GUADALEST S.A. DE C.V.", "normalized": "grupoguadalestsadecv", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "R.F.C.", "value": "GGU040603B20", "source_qualifier": "Mexico" }, "key": [ "rfc", "ggu040603b20" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-038", "company_name": "CIA. CONSTRUCTORA Y COMERCIALIZADORA DEL SUR LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "CIA. CONSTRUCTORA Y COMERCIALIZADORA DEL SUR LTDA.", "normalized": "ciaconstructoraycomercializadoradelsurltda", "matched_uids": [] }, { "query": "COSUR LTDA.", "normalized": "cosurltda", "matched_uids": [] }, { "query": "HOTEL PALACE", "normalized": "hotelpalace", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "890329758-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8903297587" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "CIA. CONSTRUCTORA Y COMERCIALIZADORA DEL SUR LTDA.", "normalized": "ciaconstructoraycomercializadoradelsurltda", "matched_uids": [] }, { "query": "COSUR LTDA.", "normalized": "cosurltda", "matched_uids": [] }, { "query": "HOTEL PALACE", "normalized": "hotelpalace", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "890329758-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8903297587" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-039", "company_name": "INTERCONTINENTAL DE FINANCIACION AEREA S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "INTERCONTINENTAL DE FINANCIACION AEREA S.A.", "normalized": "intercontinentaldefinanciacionaereasa", "matched_uids": [] }, { "query": "INTERFIAR S.A.", "normalized": "interfiarsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800043810-6", "source_qualifier": "Colombia" }, "key": [ "nit", "8000438106" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "INTERCONTINENTAL DE FINANCIACION AEREA S.A.", "normalized": "intercontinentaldefinanciacionaereasa", "matched_uids": [] }, { "query": "INTERFIAR S.A.", "normalized": "interfiarsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800043810-6", "source_qualifier": "Colombia" }, "key": [ "nit", "8000438106" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-040", "company_name": "MASA FACIL, S.A. DE C.V.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "MASA FACIL, S.A. DE C.V.", "normalized": "masafacilsadecv", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "Folio Mercantil No.", "value": "343997", "source_qualifier": "Mexico" }, "key": [ "foliomercantil", "343997" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "MASA FACIL, S.A. DE C.V.", "normalized": "masafacilsadecv", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "Folio Mercantil No.", "value": "343997", "source_qualifier": "Mexico" }, "key": [ "foliomercantil", "343997" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-041", "company_name": "GRUPO FALCON DE PANAMA, S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "GRUPO FALCON DE PANAMA, S.A.", "normalized": "grupofalcondepanamasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "RUC #", "value": "556350-1-444254-07", "source_qualifier": "Panama" }, "key": [ "ruc", "556350144425407" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "GRUPO FALCON DE PANAMA, S.A.", "normalized": "grupofalcondepanamasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "RUC #", "value": "556350-1-444254-07", "source_qualifier": "Panama" }, "key": [ "ruc", "556350144425407" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-042", "company_name": "COMERCIAL JINAN S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "COMERCIAL JINAN S.A.", "normalized": "comercialjinansa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "4151952-3", "source_qualifier": "Guatemala" }, "key": [ "nit", "41519523" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "COMERCIAL JINAN S.A.", "normalized": "comercialjinansa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "4151952-3", "source_qualifier": "Guatemala" }, "key": [ "nit", "41519523" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-043", "company_name": "AMD CO. LTD AGENCY", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "AMD CO. LTD AGENCY", "normalized": "amdcoltdagency", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "AMD CO. LTD AGENCY", "normalized": "amdcoltdagency", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-044", "company_name": "ALIMENTOS CARNICOS DE TRADICION ESPANOLA LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ALIMENTOS CARNICOS DE TRADICION ESPANOLA LTDA.", "normalized": "alimentoscarnicosdetradicionespanolaltda", "matched_uids": [] }, { "query": "ALICANTE", "normalized": "alicante", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "900229820-2", "source_qualifier": "Colombia" }, "key": [ "nit", "9002298202" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ALIMENTOS CARNICOS DE TRADICION ESPANOLA LTDA.", "normalized": "alimentoscarnicosdetradicionespanolaltda", "matched_uids": [] }, { "query": "ALICANTE", "normalized": "alicante", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "900229820-2", "source_qualifier": "Colombia" }, "key": [ "nit", "9002298202" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-045", "company_name": "HIERROS DE JERUSALEM S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "HIERROS DE JERUSALEM S.A.", "normalized": "hierrosdejerusalemsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830513468-6", "source_qualifier": "Colombia" }, "key": [ "nit", "8305134686" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "HIERROS DE JERUSALEM S.A.", "normalized": "hierrosdejerusalemsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830513468-6", "source_qualifier": "Colombia" }, "key": [ "nit", "8305134686" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-046", "company_name": "AGROESPINAL S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "AGROESPINAL S.A.", "normalized": "agroespinalsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800256233-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8002562330" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "AGROESPINAL S.A.", "normalized": "agroespinalsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800256233-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8002562330" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-047", "company_name": "CANALES VENECIA LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "CANALES VENECIA LTDA.", "normalized": "canalesvenecialtda", "matched_uids": [] }, { "query": "CANALVE LTDA.", "normalized": "canalveltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "8110469899", "source_qualifier": "Colombia" }, "key": [ "nit", "8110469899" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "CANALES VENECIA LTDA.", "normalized": "canalesvenecialtda", "matched_uids": [] }, { "query": "CANALVE LTDA.", "normalized": "canalveltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "8110469899", "source_qualifier": "Colombia" }, "key": [ "nit", "8110469899" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-048", "company_name": "GREEN ISLAND S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "GREEN ISLAND S.A.", "normalized": "greenislandsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830067456-4", "source_qualifier": "Colombia" }, "key": [ "nit", "8300674564" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "GREEN ISLAND S.A.", "normalized": "greenislandsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830067456-4", "source_qualifier": "Colombia" }, "key": [ "nit", "8300674564" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-049", "company_name": "SAAVEDRA Y CIA. S. EN C.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "SAAVEDRA Y CIA. S. EN C.", "normalized": "saavedrayciasenc", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "890332983-9", "source_qualifier": "Colombia" }, "key": [ "nit", "8903329839" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "SAAVEDRA Y CIA. S. EN C.", "normalized": "saavedrayciasenc", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "890332983-9", "source_qualifier": "Colombia" }, "key": [ "nit", "8903329839" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-050", "company_name": "SOCIEDAD SUPERDEPORTES LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "SOCIEDAD SUPERDEPORTES LTDA.", "normalized": "sociedadsuperdeportesltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "8009712337", "source_qualifier": "Colombia" }, "key": [ "nit", "8009712337" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "SOCIEDAD SUPERDEPORTES LTDA.", "normalized": "sociedadsuperdeportesltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "8009712337", "source_qualifier": "Colombia" }, "key": [ "nit", "8009712337" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-051", "company_name": "SERVIAGRICOLA CIFUENTES E.U.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "SERVIAGRICOLA CIFUENTES E.U.", "normalized": "serviagricolacifuenteseu", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "805025920-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8050259201" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "SERVIAGRICOLA CIFUENTES E.U.", "normalized": "serviagricolacifuenteseu", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "805025920-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8050259201" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-052", "company_name": "TAXI AEREO ANTIOQUENO S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "TAXI AEREO ANTIOQUENO S.A.", "normalized": "taxiaereoantioquenosa", "matched_uids": [] }, { "query": "TAN S.A.", "normalized": "tansa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "811041365-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8110413650" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "TAXI AEREO ANTIOQUENO S.A.", "normalized": "taxiaereoantioquenosa", "matched_uids": [] }, { "query": "TAN S.A.", "normalized": "tansa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "811041365-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8110413650" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-053", "company_name": "TIGRIS TRADING, INC.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "TIGRIS TRADING, INC.", "normalized": "tigristradinginc", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "TIGRIS TRADING, INC.", "normalized": "tigristradinginc", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-054", "company_name": "TRANS PACIFIC WORLD LEASING LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "TRANS PACIFIC WORLD LEASING LIMITED", "normalized": "transpacificworldleasinglimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "TRANS PACIFIC WORLD LEASING LIMITED", "normalized": "transpacificworldleasinglimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-055", "company_name": "CANO AGUDELO S EN C", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "CANO AGUDELO S EN C", "normalized": "canoagudelosenc", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "821002095-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8210020957" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "CANO AGUDELO S EN C", "normalized": "canoagudelosenc", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "821002095-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8210020957" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-056", "company_name": "ACCIRENT S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ACCIRENT S.A.", "normalized": "accirentsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830088969-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8300889690" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ACCIRENT S.A.", "normalized": "accirentsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "830088969-0", "source_qualifier": "Colombia" }, "key": [ "nit", "8300889690" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-057", "company_name": "TARIQ ABU SHANAB EST. FOR TRADE & COMMERCE", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "TARIQ ABU SHANAB EST. FOR TRADE & COMMERCE", "normalized": "tariqabushanabestfortradeandcommerce", "matched_uids": [] }, { "query": "ABU SHANAB METALS ESTABLISHMENT", "normalized": "abushanabmetalsestablishment", "matched_uids": [] }, { "query": "AMIN ABU SHANAB & SONS CO.", "normalized": "aminabushanabandsonsco", "matched_uids": [] }, { "query": "AMIN ABU SHANAB AND SONS CO.", "normalized": "aminabushanabandsonsco", "matched_uids": [] }, { "query": "SHANAB METALS ESTABLISHMENT", "normalized": "shanabmetalsestablishment", "matched_uids": [] }, { "query": "TARIQ ABU SHANAB EST.", "normalized": "tariqabushanabest", "matched_uids": [] }, { "query": "TARIQ ABU SHANAB EST. FOR TRADE AND COMMERCE", "normalized": "tariqabushanabestfortradeandcommerce", "matched_uids": [] }, { "query": "TARIQ ABU SHANAB METALS ESTABLISHMENT", "normalized": "tariqabushanabmetalsestablishment", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "TARIQ ABU SHANAB EST. FOR TRADE & COMMERCE", "normalized": "tariqabushanabestfortradeandcommerce", "matched_uids": [] }, { "query": "ABU SHANAB METALS ESTABLISHMENT", "normalized": "abushanabmetalsestablishment", "matched_uids": [] }, { "query": "AMIN ABU SHANAB & SONS CO.", "normalized": "aminabushanabandsonsco", "matched_uids": [] }, { "query": "AMIN ABU SHANAB AND SONS CO.", "normalized": "aminabushanabandsonsco", "matched_uids": [] }, { "query": "SHANAB METALS ESTABLISHMENT", "normalized": "shanabmetalsestablishment", "matched_uids": [] }, { "query": "TARIQ ABU SHANAB EST.", "normalized": "tariqabushanabest", "matched_uids": [] }, { "query": "TARIQ ABU SHANAB EST. FOR TRADE AND COMMERCE", "normalized": "tariqabushanabestfortradeandcommerce", "matched_uids": [] }, { "query": "TARIQ ABU SHANAB METALS ESTABLISHMENT", "normalized": "tariqabushanabmetalsestablishment", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-058", "company_name": "ASES DE COMPETENCIA Y CIA. S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "ASES DE COMPETENCIA Y CIA. S.A.", "normalized": "asesdecompetenciayciasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800213156-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8002131567" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "ASES DE COMPETENCIA Y CIA. S.A.", "normalized": "asesdecompetenciayciasa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800213156-7", "source_qualifier": "Colombia" }, "key": [ "nit", "8002131567" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-059", "company_name": "DURAND PROPERTIES LIMITED", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "DURAND PROPERTIES LIMITED", "normalized": "durandpropertieslimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "DURAND PROPERTIES LIMITED", "normalized": "durandpropertieslimited", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-060", "company_name": "RESTAURANTE BAR PUNTA DEL ESTE", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "RESTAURANTE BAR PUNTA DEL ESTE", "normalized": "restaurantebarpuntadeleste", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "Matricula Mercantil No", "value": "387183-1", "source_qualifier": "Colombia" }, "key": [ "matriculamercantil", "3871831" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "RESTAURANTE BAR PUNTA DEL ESTE", "normalized": "restaurantebarpuntadeleste", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "Matricula Mercantil No", "value": "387183-1", "source_qualifier": "Colombia" }, "key": [ "matriculamercantil", "3871831" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-061", "company_name": "AGROGANADERA LOS SANTOS S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "AGROGANADERA LOS SANTOS S.A.", "normalized": "agroganaderalossantossa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800215934-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8002159341" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "AGROGANADERA LOS SANTOS S.A.", "normalized": "agroganaderalossantossa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800215934-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8002159341" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-062", "company_name": "GRUPO FALCON S.A.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "GRUPO FALCON S.A.", "normalized": "grupofalconsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800214711-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8002147111" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "GRUPO FALCON S.A.", "normalized": "grupofalconsa", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800214711-1", "source_qualifier": "Colombia" }, "key": [ "nit", "8002147111" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-063", "company_name": "LARGO LEASING LTD.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "LARGO LEASING LTD.", "normalized": "largoleasingltd", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "LARGO LEASING LTD.", "normalized": "largoleasingltd", "matched_uids": [] } ], "corporate_identifiers": [], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-064", "company_name": "BOSQUE DE SANTA TERESITA LTDA.", "expected_current_uids": [], "official_lookup": { "names": [ { "query": "BOSQUE DE SANTA TERESITA LTDA.", "normalized": "bosquedesantateresitaltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800117606-9", "source_qualifier": "Colombia" }, "key": [ "nit", "8001176069" ], "matched_uids": [] } ], "all_matched_uids": [] }, "product_lookup": { "names": [ { "query": "BOSQUE DE SANTA TERESITA LTDA.", "normalized": "bosquedesantateresitaltda", "matched_uids": [] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800117606-9", "source_qualifier": "Colombia" }, "key": [ "nit", "8001176069" ], "matched_uids": [] } ], "all_matched_uids": [] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-065", "company_name": "CREDISA S.A.", "expected_current_uids": [ "7583" ], "official_lookup": { "names": [ { "query": "CREDISA S.A.", "normalized": "credisasa", "matched_uids": [ "7583" ] }, { "query": "COMERCIALIZADORA AUTOMOTRIZ S.A.", "normalized": "comercializadoraautomotrizsa", "matched_uids": [ "7583" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800065773-6", "country": "Colombia" }, "key": [ "nit", "8000657736" ], "matched_uids": [ "7583" ] } ], "all_matched_uids": [ "7583" ] }, "product_lookup": { "names": [ { "query": "CREDISA S.A.", "normalized": "credisasa", "matched_uids": [ "7583" ] }, { "query": "COMERCIALIZADORA AUTOMOTRIZ S.A.", "normalized": "comercializadoraautomotrizsa", "matched_uids": [ "7583" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800065773-6", "country": "Colombia" }, "key": [ "nit", "8000657736" ], "matched_uids": [ "7583" ] } ], "all_matched_uids": [ "7583" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-066", "company_name": "MIRA E.U.", "expected_current_uids": [ "7593" ], "official_lookup": { "names": [ { "query": "MIRA E.U.", "normalized": "miraeu", "matched_uids": [ "7593" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "805009267-0", "country": "Colombia" }, "key": [ "nit", "8050092670" ], "matched_uids": [ "7593" ] } ], "all_matched_uids": [ "7593" ] }, "product_lookup": { "names": [ { "query": "MIRA E.U.", "normalized": "miraeu", "matched_uids": [ "7593" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "805009267-0", "country": "Colombia" }, "key": [ "nit", "8050092670" ], "matched_uids": [ "7593" ] } ], "all_matched_uids": [ "7593" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-067", "company_name": "CIA. MINERA DAPA S.A.", "expected_current_uids": [ "7579" ], "official_lookup": { "names": [ { "query": "CIA. MINERA DAPA S.A.", "normalized": "ciamineradapasa", "matched_uids": [ "7579" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800181373-1", "country": "Colombia" }, "key": [ "nit", "8001813731" ], "matched_uids": [ "7579" ] } ], "all_matched_uids": [ "7579" ] }, "product_lookup": { "names": [ { "query": "CIA. MINERA DAPA S.A.", "normalized": "ciamineradapasa", "matched_uids": [ "7579" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800181373-1", "country": "Colombia" }, "key": [ "nit", "8001813731" ], "matched_uids": [ "7579" ] } ], "all_matched_uids": [ "7579" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-068", "company_name": "INMOBILIARIA EL ESCORIAL LTDA.", "expected_current_uids": [ "7288" ], "official_lookup": { "names": [ { "query": "INMOBILIARIA EL ESCORIAL LTDA.", "normalized": "inmobiliariaelescorialltda", "matched_uids": [ "7288" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146869-2", "country": "Colombia" }, "key": [ "nit", "8001468692" ], "matched_uids": [ "7288" ] } ], "all_matched_uids": [ "7288" ] }, "product_lookup": { "names": [ { "query": "INMOBILIARIA EL ESCORIAL LTDA.", "normalized": "inmobiliariaelescorialltda", "matched_uids": [ "7288" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146869-2", "country": "Colombia" }, "key": [ "nit", "8001468692" ], "matched_uids": [ "7288" ] } ], "all_matched_uids": [ "7288" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-069", "company_name": "AGRICOLA DOIMA DEL NORTE DEL VALLE LTDA.", "expected_current_uids": [ "7283" ], "official_lookup": { "names": [ { "query": "AGRICOLA DOIMA DEL NORTE DEL VALLE LTDA.", "normalized": "agricoladoimadelnortedelvalleltda", "matched_uids": [ "7283" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800144713-3", "country": "Colombia" }, "key": [ "nit", "8001447133" ], "matched_uids": [ "7283" ] } ], "all_matched_uids": [ "7283" ] }, "product_lookup": { "names": [ { "query": "AGRICOLA DOIMA DEL NORTE DEL VALLE LTDA.", "normalized": "agricoladoimadelnortedelvalleltda", "matched_uids": [ "7283" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800144713-3", "country": "Colombia" }, "key": [ "nit", "8001447133" ], "matched_uids": [ "7283" ] } ], "all_matched_uids": [ "7283" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-070", "company_name": "INMOBILIARIA LINARES LTDA.", "expected_current_uids": [ "7289" ], "official_lookup": { "names": [ { "query": "INMOBILIARIA LINARES LTDA.", "normalized": "inmobiliarialinaresltda", "matched_uids": [ "7289" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146860-7", "country": "Colombia" }, "key": [ "nit", "8001468607" ], "matched_uids": [ "7289" ] } ], "all_matched_uids": [ "7289" ] }, "product_lookup": { "names": [ { "query": "INMOBILIARIA LINARES LTDA.", "normalized": "inmobiliarialinaresltda", "matched_uids": [ "7289" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146860-7", "country": "Colombia" }, "key": [ "nit", "8001468607" ], "matched_uids": [ "7289" ] } ], "all_matched_uids": [ "7289" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-071", "company_name": "AGROPECUARIA MIRALINDO S.A.", "expected_current_uids": [ "7284" ], "official_lookup": { "names": [ { "query": "AGROPECUARIA MIRALINDO S.A.", "normalized": "agropecuariamiralindosa", "matched_uids": [ "7284" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "836000446-4", "country": "Colombia" }, "key": [ "nit", "8360004464" ], "matched_uids": [ "7284" ] } ], "all_matched_uids": [ "7284" ] }, "product_lookup": { "names": [ { "query": "AGROPECUARIA MIRALINDO S.A.", "normalized": "agropecuariamiralindosa", "matched_uids": [ "7284" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "836000446-4", "country": "Colombia" }, "key": [ "nit", "8360004464" ], "matched_uids": [ "7284" ] } ], "all_matched_uids": [ "7284" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-072", "company_name": "BANANERA AGRICOLA S.A.", "expected_current_uids": [ "7576" ], "official_lookup": { "names": [ { "query": "BANANERA AGRICOLA S.A.", "normalized": "bananeraagricolasa", "matched_uids": [ "7576" ] }, { "query": "BANAGRICOLA S.A.", "normalized": "banagricolasa", "matched_uids": [ "7576" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800142651-6", "country": "Colombia" }, "key": [ "nit", "8001426516" ], "matched_uids": [ "7576" ] } ], "all_matched_uids": [ "7576" ] }, "product_lookup": { "names": [ { "query": "BANANERA AGRICOLA S.A.", "normalized": "bananeraagricolasa", "matched_uids": [ "7576" ] }, { "query": "BANAGRICOLA S.A.", "normalized": "banagricolasa", "matched_uids": [ "7576" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800142651-6", "country": "Colombia" }, "key": [ "nit", "8001426516" ], "matched_uids": [ "7576" ] } ], "all_matched_uids": [ "7576" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-073", "company_name": "GANADERIA EL VERGEL LTDA.", "expected_current_uids": [ "7286" ], "official_lookup": { "names": [ { "query": "GANADERIA EL VERGEL LTDA.", "normalized": "ganaderiaelvergelltda", "matched_uids": [ "7286" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146295-5", "country": "Colombia" }, "key": [ "nit", "8001462955" ], "matched_uids": [ "7286" ] } ], "all_matched_uids": [ "7286" ] }, "product_lookup": { "names": [ { "query": "GANADERIA EL VERGEL LTDA.", "normalized": "ganaderiaelvergelltda", "matched_uids": [ "7286" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146295-5", "country": "Colombia" }, "key": [ "nit", "8001462955" ], "matched_uids": [ "7286" ] } ], "all_matched_uids": [ "7286" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-074", "company_name": "CRIADERO LA LUISA E.U.", "expected_current_uids": [ "7584" ], "official_lookup": { "names": [ { "query": "CRIADERO LA LUISA E.U.", "normalized": "criaderolaluisaeu", "matched_uids": [ "7584" ] }, { "query": "INDUSTRIA AGROPECUARIA SANTA ELENA LTDA.", "normalized": "industriaagropecuariasantaelenaltda", "matched_uids": [ "7584" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "860503330-5", "country": "Colombia" }, "key": [ "nit", "8605033305" ], "matched_uids": [ "7584" ] } ], "all_matched_uids": [ "7584" ] }, "product_lookup": { "names": [ { "query": "CRIADERO LA LUISA E.U.", "normalized": "criaderolaluisaeu", "matched_uids": [ "7584" ] }, { "query": "INDUSTRIA AGROPECUARIA SANTA ELENA LTDA.", "normalized": "industriaagropecuariasantaelenaltda", "matched_uids": [ "7584" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "860503330-5", "country": "Colombia" }, "key": [ "nit", "8605033305" ], "matched_uids": [ "7584" ] } ], "all_matched_uids": [ "7584" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-075", "company_name": "GOMEZ MARIN LTDA.", "expected_current_uids": [ "10003" ], "official_lookup": { "names": [ { "query": "GOMEZ MARIN LTDA.", "normalized": "gomezmarinltda", "matched_uids": [ "10003" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800102465-1", "country": "Colombia" }, "key": [ "nit", "8001024651" ], "matched_uids": [ "10003" ] } ], "all_matched_uids": [ "10003" ] }, "product_lookup": { "names": [ { "query": "GOMEZ MARIN LTDA.", "normalized": "gomezmarinltda", "matched_uids": [ "10003" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800102465-1", "country": "Colombia" }, "key": [ "nit", "8001024651" ], "matched_uids": [ "10003" ] } ], "all_matched_uids": [ "10003" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-076", "company_name": "INMOBILIARIA PASADENA LTDA.", "expected_current_uids": [ "7290" ], "official_lookup": { "names": [ { "query": "INMOBILIARIA PASADENA LTDA.", "normalized": "inmobiliariapasadenaltda", "matched_uids": [ "7290" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146861-4", "country": "Colombia" }, "key": [ "nit", "8001468614" ], "matched_uids": [ "7290" ] } ], "all_matched_uids": [ "7290" ] }, "product_lookup": { "names": [ { "query": "INMOBILIARIA PASADENA LTDA.", "normalized": "inmobiliariapasadenaltda", "matched_uids": [ "7290" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800146861-4", "country": "Colombia" }, "key": [ "nit", "8001468614" ], "matched_uids": [ "7290" ] } ], "all_matched_uids": [ "7290" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-077", "company_name": "LADRILLERA LA CANDELARIA LTDA.", "expected_current_uids": [ "7292" ], "official_lookup": { "names": [ { "query": "LADRILLERA LA CANDELARIA LTDA.", "normalized": "ladrilleralacandelarialtda", "matched_uids": [ "7292" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800119741-4", "country": "Colombia" }, "key": [ "nit", "8001197414" ], "matched_uids": [ "7292" ] } ], "all_matched_uids": [ "7292" ] }, "product_lookup": { "names": [ { "query": "LADRILLERA LA CANDELARIA LTDA.", "normalized": "ladrilleralacandelarialtda", "matched_uids": [ "7292" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800119741-4", "country": "Colombia" }, "key": [ "nit", "8001197414" ], "matched_uids": [ "7292" ] } ], "all_matched_uids": [ "7292" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-078", "company_name": "VISCAYA LTDA.", "expected_current_uids": [ "7295" ], "official_lookup": { "names": [ { "query": "VISCAYA LTDA.", "normalized": "viscayaltda", "matched_uids": [ "7295" ] }, { "query": "VIZCAYA LTDA.", "normalized": "vizcayaltda", "matched_uids": [ "7295" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800054357-8", "country": "Colombia" }, "key": [ "nit", "8000543578" ], "matched_uids": [ "7295" ] } ], "all_matched_uids": [ "7295" ] }, "product_lookup": { "names": [ { "query": "VISCAYA LTDA.", "normalized": "viscayaltda", "matched_uids": [ "7295" ] }, { "query": "VIZCAYA LTDA.", "normalized": "vizcayaltda", "matched_uids": [ "7295" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800054357-8", "country": "Colombia" }, "key": [ "nit", "8000543578" ], "matched_uids": [ "7295" ] } ], "all_matched_uids": [ "7295" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-079", "company_name": "GAVIOTAS LTDA.", "expected_current_uids": [ "10004" ], "official_lookup": { "names": [ { "query": "GAVIOTAS LTDA.", "normalized": "gaviotasltda", "matched_uids": [ "10004" ] }, { "query": "MOTEL CAMPO AMOR", "normalized": "motelcampoamor", "matched_uids": [ "10004" ] }, { "query": "HOTEL SIN PECADOS", "normalized": "hotelsinpecados", "matched_uids": [ "10004" ] }, { "query": "HOTEL SIN PK2", "normalized": "hotelsinpk2", "matched_uids": [ "10004" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800032092-7", "country": "Colombia" }, "key": [ "nit", "8000320927" ], "matched_uids": [ "10004" ] } ], "all_matched_uids": [ "10004" ] }, "product_lookup": { "names": [ { "query": "GAVIOTAS LTDA.", "normalized": "gaviotasltda", "matched_uids": [ "10004" ] }, { "query": "MOTEL CAMPO AMOR", "normalized": "motelcampoamor", "matched_uids": [ "10004" ] }, { "query": "HOTEL SIN PECADOS", "normalized": "hotelsinpecados", "matched_uids": [ "10004" ] }, { "query": "HOTEL SIN PK2", "normalized": "hotelsinpk2", "matched_uids": [ "10004" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800032092-7", "country": "Colombia" }, "key": [ "nit", "8000320927" ], "matched_uids": [ "10004" ] } ], "all_matched_uids": [ "10004" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-080", "company_name": "TODOBOLSAS Y COLSOBRES", "expected_current_uids": [ "7631" ], "official_lookup": { "names": [ { "query": "TODOBOLSAS Y COLSOBRES", "normalized": "todobolsasycolsobres", "matched_uids": [ "7631" ] }, { "query": "RODRIGUEZ CARRENO LTDA. TODO BOLSAS Y COLSOBRES", "normalized": "rodriguezcarrenoltdatodobolsasycolsobres", "matched_uids": [ "7631" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "860053774-1", "country": "Colombia" }, "key": [ "nit", "8600537741" ], "matched_uids": [ "7631" ] } ], "all_matched_uids": [ "7631" ] }, "product_lookup": { "names": [ { "query": "TODOBOLSAS Y COLSOBRES", "normalized": "todobolsasycolsobres", "matched_uids": [ "7631" ] }, { "query": "RODRIGUEZ CARRENO LTDA. TODO BOLSAS Y COLSOBRES", "normalized": "rodriguezcarrenoltdatodobolsasycolsobres", "matched_uids": [ "7631" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "860053774-1", "country": "Colombia" }, "key": [ "nit", "8600537741" ], "matched_uids": [ "7631" ] } ], "all_matched_uids": [ "7631" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-081", "company_name": "COMERCIALIZADORA DE CAFE DEL OCCIDENTE CODECAFE LTDA.", "expected_current_uids": [ "10002" ], "official_lookup": { "names": [ { "query": "COMERCIALIZADORA DE CAFE DEL OCCIDENTE CODECAFE LTDA.", "normalized": "comercializadoradecafedeloccidentecodecafeltda", "matched_uids": [ "10002" ] }, { "query": "CODECAFE", "normalized": "codecafe", "matched_uids": [ "10002" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "816004106-0", "country": "Colombia" }, "key": [ "nit", "8160041060" ], "matched_uids": [ "10002" ] } ], "all_matched_uids": [ "10002" ] }, "product_lookup": { "names": [ { "query": "COMERCIALIZADORA DE CAFE DEL OCCIDENTE CODECAFE LTDA.", "normalized": "comercializadoradecafedeloccidentecodecafeltda", "matched_uids": [ "10002" ] }, { "query": "CODECAFE", "normalized": "codecafe", "matched_uids": [ "10002" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "816004106-0", "country": "Colombia" }, "key": [ "nit", "8160041060" ], "matched_uids": [ "10002" ] } ], "all_matched_uids": [ "10002" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-082", "company_name": "COMPANIA DE FOMENTO MERCANTIL S.A.", "expected_current_uids": [ "7580" ], "official_lookup": { "names": [ { "query": "COMPANIA DE FOMENTO MERCANTIL S.A.", "normalized": "companiadefomentomercantilsa", "matched_uids": [ "7580" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800124589-0", "country": "Colombia" }, "key": [ "nit", "8001245890" ], "matched_uids": [ "7580" ] } ], "all_matched_uids": [ "7580" ] }, "product_lookup": { "names": [ { "query": "COMPANIA DE FOMENTO MERCANTIL S.A.", "normalized": "companiadefomentomercantilsa", "matched_uids": [ "7580" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800124589-0", "country": "Colombia" }, "key": [ "nit", "8001245890" ], "matched_uids": [ "7580" ] } ], "all_matched_uids": [ "7580" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-083", "company_name": "GRAN MUELLE S.A.", "expected_current_uids": [ "7590" ], "official_lookup": { "names": [ { "query": "GRAN MUELLE S.A.", "normalized": "granmuellesa", "matched_uids": [ "7590" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800173054-1", "country": "Colombia" }, "key": [ "nit", "8001730541" ], "matched_uids": [ "7590" ] } ], "all_matched_uids": [ "7590" ] }, "product_lookup": { "names": [ { "query": "GRAN MUELLE S.A.", "normalized": "granmuellesa", "matched_uids": [ "7590" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800173054-1", "country": "Colombia" }, "key": [ "nit", "8001730541" ], "matched_uids": [ "7590" ] } ], "all_matched_uids": [ "7590" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-084", "company_name": "UNIDAS S.A.", "expected_current_uids": [ "7600" ], "official_lookup": { "names": [ { "query": "UNIDAS S.A.", "normalized": "unidassa", "matched_uids": [ "7600" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800080688-0", "country": "Colombia" }, "key": [ "nit", "8000806880" ], "matched_uids": [ "7600" ] } ], "all_matched_uids": [ "7600" ] }, "product_lookup": { "names": [ { "query": "UNIDAS S.A.", "normalized": "unidassa", "matched_uids": [ "7600" ] } ], "corporate_identifiers": [ { "query": { "type": "NIT #", "value": "800080688-0", "country": "Colombia" }, "key": [ "nit", "8000806880" ], "matched_uids": [ "7600" ] } ], "all_matched_uids": [ "7600" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-085", "company_name": "AL WASEL AND BABEL GENERAL TRADING LLC", "expected_current_uids": [ "8210" ], "official_lookup": { "names": [ { "query": "AL WASEL AND BABEL GENERAL TRADING LLC", "normalized": "alwaselandbabelgeneraltradingllc", "matched_uids": [ "8210" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8210" ] }, "product_lookup": { "names": [ { "query": "AL WASEL AND BABEL GENERAL TRADING LLC", "normalized": "alwaselandbabelgeneraltradingllc", "matched_uids": [ "8210" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8210" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-086", "company_name": "SES INTERNATIONAL CORP", "expected_current_uids": [ "8882" ], "official_lookup": { "names": [ { "query": "SES INTERNATIONAL CORP", "normalized": "sesinternationalcorp", "matched_uids": [ "8882" ] }, { "query": "SES GROUP", "normalized": "sesgroup", "matched_uids": [ "8882" ] }, { "query": "SES AUTOMOBILE", "normalized": "sesautomobile", "matched_uids": [ "8882" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8882" ] }, "product_lookup": { "names": [ { "query": "SES INTERNATIONAL CORP", "normalized": "sesinternationalcorp", "matched_uids": [ "8882" ] }, { "query": "SES GROUP", "normalized": "sesgroup", "matched_uids": [ "8882" ] }, { "query": "SES AUTOMOBILE", "normalized": "sesautomobile", "matched_uids": [ "8882" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8882" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-087", "company_name": "TECHNOLOGY AND DEVELOPMENT GROUP LTD", "expected_current_uids": [ "8408" ], "official_lookup": { "names": [ { "query": "TECHNOLOGY AND DEVELOPMENT GROUP LTD", "normalized": "technologyanddevelopmentgroupltd", "matched_uids": [ "8408" ] }, { "query": "T.D.G.", "normalized": "tdg", "matched_uids": [ "8408" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8408" ] }, "product_lookup": { "names": [ { "query": "TECHNOLOGY AND DEVELOPMENT GROUP LTD", "normalized": "technologyanddevelopmentgroupltd", "matched_uids": [ "8408" ] }, { "query": "T.D.G.", "normalized": "tdg", "matched_uids": [ "8408" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8408" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-088", "company_name": "AVIATRANS ANSTALT", "expected_current_uids": [ "8213" ], "official_lookup": { "names": [ { "query": "AVIATRANS ANSTALT", "normalized": "aviatransanstalt", "matched_uids": [ "8213" ] }, { "query": "AVIATRANS ESTABLISHMENT", "normalized": "aviatransestablishment", "matched_uids": [ "8213" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8213" ] }, "product_lookup": { "names": [ { "query": "AVIATRANS ANSTALT", "normalized": "aviatransanstalt", "matched_uids": [ "8213" ] }, { "query": "AVIATRANS ESTABLISHMENT", "normalized": "aviatransestablishment", "matched_uids": [ "8213" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8213" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-089", "company_name": "AL-HUDA STATE COMPANY FOR RELIGIOUS TOURISM", "expected_current_uids": [ "8212" ], "official_lookup": { "names": [ { "query": "AL-HUDA STATE COMPANY FOR RELIGIOUS TOURISM", "normalized": "alhudastatecompanyforreligioustourism", "matched_uids": [ "8212" ] }, { "query": "AL-HODA FOR RELIGIOUS TOURISM COMPANY", "normalized": "alhodaforreligioustourismcompany", "matched_uids": [ "8212" ] }, { "query": "AL-HODA STATE COMPANY FOR RELIGIOUS TOURISM", "normalized": "alhodastatecompanyforreligioustourism", "matched_uids": [ "8212" ] }, { "query": "AL-HUDA FOR RELIGIOUS TOURISM COMPANY", "normalized": "alhudaforreligioustourismcompany", "matched_uids": [ "8212" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8212" ] }, "product_lookup": { "names": [ { "query": "AL-HUDA STATE COMPANY FOR RELIGIOUS TOURISM", "normalized": "alhudastatecompanyforreligioustourism", "matched_uids": [ "8212" ] }, { "query": "AL-HODA FOR RELIGIOUS TOURISM COMPANY", "normalized": "alhodaforreligioustourismcompany", "matched_uids": [ "8212" ] }, { "query": "AL-HODA STATE COMPANY FOR RELIGIOUS TOURISM", "normalized": "alhodastatecompanyforreligioustourism", "matched_uids": [ "8212" ] }, { "query": "AL-HUDA FOR RELIGIOUS TOURISM COMPANY", "normalized": "alhudaforreligioustourismcompany", "matched_uids": [ "8212" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8212" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-090", "company_name": "T.M.G. ENGINEERING LIMITED", "expected_current_uids": [ "8405" ], "official_lookup": { "names": [ { "query": "T.M.G. ENGINEERING LIMITED", "normalized": "tmgengineeringlimited", "matched_uids": [ "8405" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8405" ] }, "product_lookup": { "names": [ { "query": "T.M.G. ENGINEERING LIMITED", "normalized": "tmgengineeringlimited", "matched_uids": [ "8405" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8405" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-091", "company_name": "LOGARCHEO S.A.", "expected_current_uids": [ "8214" ], "official_lookup": { "names": [ { "query": "LOGARCHEO S.A.", "normalized": "logarcheosa", "matched_uids": [ "8214" ] }, { "query": "LOGARCHEO AG", "normalized": "logarcheoag", "matched_uids": [ "8214" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8214" ] }, "product_lookup": { "names": [ { "query": "LOGARCHEO S.A.", "normalized": "logarcheosa", "matched_uids": [ "8214" ] }, { "query": "LOGARCHEO AG", "normalized": "logarcheoag", "matched_uids": [ "8214" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8214" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-092", "company_name": "AL-ARABI TRADING COMPANY LIMITED", "expected_current_uids": [ "8314" ], "official_lookup": { "names": [ { "query": "AL-ARABI TRADING COMPANY LIMITED", "normalized": "alarabitradingcompanylimited", "matched_uids": [ "8314" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8314" ] }, "product_lookup": { "names": [ { "query": "AL-ARABI TRADING COMPANY LIMITED", "normalized": "alarabitradingcompanylimited", "matched_uids": [ "8314" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "8314" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-093", "company_name": "NEW AHMADI LTD.", "expected_current_uids": [ "13128" ], "official_lookup": { "names": [ { "query": "NEW AHMADI LTD.", "normalized": "newahmadiltd", "matched_uids": [ "13128" ] }, { "query": "NEW AHMADI COMPANY LTD", "normalized": "newahmadicompanyltd", "matched_uids": [ "13128" ] }, { "query": "NEW AHMADY LTD", "normalized": "newahmadyltd", "matched_uids": [ "13128" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "13128" ] }, "product_lookup": { "names": [ { "query": "NEW AHMADI LTD.", "normalized": "newahmadiltd", "matched_uids": [ "13128" ] }, { "query": "NEW AHMADI COMPANY LTD", "normalized": "newahmadicompanyltd", "matched_uids": [ "13128" ] }, { "query": "NEW AHMADY LTD", "normalized": "newahmadyltd", "matched_uids": [ "13128" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "13128" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-094", "company_name": "INMOBILIARIA TIJUANA COSTA S.A. DE C.V.", "expected_current_uids": [ "9975" ], "official_lookup": { "names": [ { "query": "INMOBILIARIA TIJUANA COSTA S.A. DE C.V.", "normalized": "inmobiliariatijuanacostasadecv", "matched_uids": [ "9975" ] } ], "corporate_identifiers": [ { "query": { "type": "R.F.C.", "value": "ITC-910503", "country": "Mexico" }, "key": [ "rfc", "itc910503" ], "matched_uids": [ "9975" ] } ], "all_matched_uids": [ "9975" ] }, "product_lookup": { "names": [ { "query": "INMOBILIARIA TIJUANA COSTA S.A. DE C.V.", "normalized": "inmobiliariatijuanacostasadecv", "matched_uids": [ "9975" ] } ], "corporate_identifiers": [ { "query": { "type": "R.F.C.", "value": "ITC-910503", "country": "Mexico" }, "key": [ "rfc", "itc910503" ], "matched_uids": [ "9975" ] } ], "all_matched_uids": [ "9975" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-095", "company_name": "ACCESOS ELECTRONICOS S.A.DE C.V.", "expected_current_uids": [ "7151" ], "official_lookup": { "names": [ { "query": "ACCESOS ELECTRONICOS S.A.DE C.V.", "normalized": "accesoselectronicossadecv", "matched_uids": [ "7151" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "7151" ] }, "product_lookup": { "names": [ { "query": "ACCESOS ELECTRONICOS S.A.DE C.V.", "normalized": "accesoselectronicossadecv", "matched_uids": [ "7151" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "7151" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true }, { "case_id": "api-case-096", "company_name": "SIAM NICE COMPANY LTD.", "expected_current_uids": [ "9541" ], "official_lookup": { "names": [ { "query": "SIAM NICE COMPANY LTD.", "normalized": "siamnicecompanyltd", "matched_uids": [ "9541" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "9541" ] }, "product_lookup": { "names": [ { "query": "SIAM NICE COMPANY LTD.", "normalized": "siamnicecompanyltd", "matched_uids": [ "9541" ] } ], "corporate_identifiers": [], "all_matched_uids": [ "9541" ] }, "official_matches_expected": true, "product_matches_expected": true, "every_selected_name_identifier_lookup_agrees": true, "selected_current_record_common_fields_equal": true } ] }, "product_initial_preparation_observation": "At17:10:51 UTC production still exposed the Oct5 source. The end read matched Oct8 and recorded retrieval at17:11:22 UTC. Initial selected-panel agreement did not imply whole-source parity; no production refresh was triggered by the researchers.", "analysis_input_sha256": "b5bd6215bddbf68fc4978745a99bfb3e7b0b08a37999a249768a54eef4cf832f", "reviews_sha256": { "M": "6c5cbdcc9ae2e9c00549b90e7f6dedee5683eac5c80e016ccc79ec68fe1f0db2", "W": "81a619e660980e39fc801f90f8ce4379ee588daf7648c6f7f08cbe89a4e045c8", "E": "0d454fe8eee9aa26f3f04f28d14aca2f88132c9d061e800bad055663409d9516" }, "limitations": "Selected96-company panel; two same-provider API models; repeated company answers; three removal events; supplied-evidence advantage; no hosted screening accuracy claim; complete historical source archives available by request." } }