How to build custom LLM-as-a-judge metrics: evaluation criteria, judge prompts, scoring types, calibration against human labels, and bias mitigation.