Adversarial self-audit

The self-audit/v1 harness measures whether Prosecutor falsely speaks on deliberately correct, behavior-preserving refactors. The initial checked-in corpus contains helper rename, extract-method, and loop-to-comprehension pairs. Each pair includes a behavior contract that must pass on both states before the item enters the denominator.

From engine/:

python3 -m exhibit_a.cli self-audit ../fixtures/refactor_corpus --offline
python3 -m exhibit_a.cli self-audit ../fixtures/refactor_corpus --docker

The first command is a fast plumbing check with the deterministic stub. The second uses the live Codex generator and hardened executor. Reports default to .exhibit-a/research/self-audit and include every raw Case, invalid-corpus reason, and per-category result. They are private research records; claims and model output may be sensitive.

Any VERIFIED Case on a validated behavior-preserving pair is a false conviction, regardless of its later intent label. The report shows the observed rate and a two-sided Wilson 95% interval. Small clean corpora retain wide upper bounds: 0/3 is not evidence of perfect precision.

The web dashboard is also private by default. Set EXHIBIT_A_RESEARCH_DASHBOARD=1 in a trusted local environment to show the latest report. EXHIBIT_A_SELF_AUDIT_DIR can point it at a different report directory.

This corpus is a foundation, not a publishable benchmark result. Expand categories, languages, repository shapes, and sample size; pre-register the analysis; and have humans verify that every transform truly preserves its declared public behavior before making false-conviction claims.