Judge prompt for code-review evals: grades whether a review's issue severities (blocker/major/minor/nit) are well-calibrated to real impact and grounded in the diff, returning a 0-1 score. Judges calibration and groundedness only, not completeness.
The LLM judge behind the code-review eval. Given a diff and the reviewer's reported issues, it scores how well the severities are calibrated to real impact and how well each issue is grounded in the shown lines — returning { reasoning, score } with score in [0, 1].
It deliberately judges calibration and groundedness only, not completeness: it does not penalize a review for issues it would personally have raised. Deterministic checks (planted-bug detection, clean-diff approval) live in the eval itself; this judge covers the part CEL can't — is a "blocker" really a blocker?
Referenced as @sufleur/[email protected] from the code-review eval's judges block; the eval asserts judge.severity.score > 0.7.