What problem does it solve? Research pipelines often promote results based on unverified evidence, leading to invalid claims, leaked test data, or non-reproducible metrics. This Skill audits confirmed evidence before promotion so that only valid, reproducible results support bounded claims. ## Core Features & Use Cases - Deterministic Evidence Audit: Inspects baseline and candidate commits, configs, data splits, preprocessing, seeds, commands, and raw result paths to verify metric recomputation and aggregation. - Leakage and Validity Checks: Detects data leakage, accidental test selection, missing cases, and failed runs, and verifies the implementation matches the adaptation plan. - Structured Verdict Recording: Writes a promotion audit report with one of three verdicts (audit_pass, audit_fix, audit_fail) plus full provenance of commits, configs, seeds, and commands. - Use Case: Before promoting a candidate model that beat a baseline, run this audit to confirm the comparison is valid, the full-budget result supports the bounded claim, and optionally gather an independent reviewer verdict. ## Quick Start Audit the confirmed experiment evidence and write a promotion verdict to research/PROMOTION_AUDIT.md before promoting the candidate.
