Your offline regression gate uses an LLM judge called through a provider alias that always points to the newest model. With no prompt or model change on your side, scores on the unchanged 150-case golden set jumped 4 points overnight. That week the gate approved a change that later regressed in production. What is the correct fix?

1. A Rerun every eval three times and average the scores, because a 4-point jump on a fixed golden set is ordinary sampling noise from the judge.
2. B Widen the gate threshold to a 10-point drop so that judge-side swings of a few points can no longer block or approve a change.
3. C Replace the LLM judge with ROUGE against reference answers, since n-gram overlap metrics are deterministic and cannot drift between runs.
4. D Pin the judge to a fixed model snapshot, and recheck its agreement with the human-labeled calibration set before any judge upgrade.