Don't trust the number. Re-run it. Everything below is committed and frozen.

The kit

git clone github.com/jourdanlabs/benchmarks
node benchmarks/correctness/score.js      # correctness scorer, LINE_TOLERANCE=5
node benchmarks/correctness/map-codeql.js # incumbent mapper, sink-faithful
npm run test:family                       # engine family, must stay green

• Ground truths are frozen. ground-truth-58/go/java.json, correctness GTs — read-only. A re-run that edits the answer key is not a re-run.
• Findings are dual-run byte-identical. Every Vantage row was produced twice with identical bytes before filing.
• Incumbents run local and pinned. CodeQL quality packs, SonarQube Community 26.8.0.126808. Versions in the recipes, not floating.
• GT-first or it doesn't ship. Any board whose ground truth was written after the findings is refused on sight — see the 0.9315 on the Misses page.

What a receipt looks like

Finding → deterministic check → sealed receipt → you re-run it. Reproduces: proven and receipted. Cannot reproduce: refused, never claimed. No model in the verdict path.