Calibrating an LLM-as-Judge Pipeline

You've built an LLM-as-judge pipeline that scores RAG answers on correctness and faithfulness at scale, and it reports correctness = 0.92 across your gold set. A skeptical staff engineer asks: "why should I believe that number?" Explain how you would calibrate the judge and defend the number, including at least two specific biases you'd check for, and describe a concrete experiment for one of them (what you'd change, what you'd measure, what result would indicate a problem).