LLM-as-judge

Reduce bias

• blind the model (remove identifiers)
• use pairwise comparisons
• multiple judges
• calibrate with human labels

Store

Judge prompts + model version.