Gemini 3.7 Flash

Confidence

95% bootstrap interval on recall (2000 resamples over the 30 PRs):6.0–18.3points. This measures sampling variance from which PRs are in the set.

Judge noise (same submission, 3 independent re-scores): recall11.6 ± 0.0pts (n=3). This is the judge alone — the exact same findings, scored again.

Neither measures model run-to-run variance — re-running the review agent itself, not just the judge. One pass per entry.

Run configuration

Severity mix

What the model called its own findings — not recall by severity, goldens aren't severity-tagged.

Category mix

Only bug/performance/security are consistent across models — the rest is free text, bucketed as other.