LLM-as-a-judge uses a separate model to score an output. Treat its score as one quality measurement, not ground truth. section 9.5.3
Judge scores need calibration and safeguards:
- Compare scores with human evaluations; judges can favor outputs from their own model family. section 9.5.3
- Record criterion-level findings rather than relying on one unexplained score; verify findings against evidence. section 7.2.4




















