SanctionsKit AI study: reproduction package, version 1

This archive contains only public corporate research inputs and sanitized API captures. Folder names mirror the research scripts; a directory named private here contains distributable study evidence, never credentials or customer data. Complete original government downloads are retained separately; see methodology for that availability limitation.

From this directory run:
python3 scripts/analyze_api_study.py --root . --reviews private/api-review-M.json --reviews private/api-review-W.json --reviews private/api-review-E.json --final-run

This uses only Python standard-library modules and no API key or network. It recomputes first-attempt scores and company-level comparisons. The analyzer deliberately labels its output preliminary; publication approval is a separate documented step. Compare the recomputed totals and by_model_arm_cohort with results-v1.json. Exact questions and request settings are in methods/api-frozen/schedule.json. All visible answer text is preserved. Provider/account IDs and reasoning ciphertext were removed; transformation-manifest.json connects each sanitized capture to its archived original hash. The internal commitment is not a public preregistration.

The frozen frame, cohort and original build_api_cohort.py show the selection inputs and ranking implementation. Rebuilding the source frame or packets additionally requires the archived original government files and parser dependencies, available on request; the included selection script is for inspection, not a standalone end-to-end rebuild. The commitment lists all original frozen inputs, including government archives not included in this package.

SanctionsKit produced and funded the research. It is not external peer review, a population accuracy estimate, legal clearance, or a test of the hosted SanctionsKit screening API.
