Before an LLM judge blocks releases, what is the most useful calibration step?

Before an LLM judge blocks releases, what is the most useful calibration step?

1. Compare its criterion-level verdicts with blinded human labels on clear and boundary cases, then inspect false passes and false failures
2. Raise the judge temperature until its explanations become more varied
3. Give the judge the candidate model name so it can account for model quality
4. Accept high agreement between two copies of the same judge prompt as ground truth