Measuring hallucination rate without expensive human eval

Hallucination rate is the fraction of outputs that contain content not supported by ground truth (or context, in RAG). Measurement options from cheap to expensive: (1) LLM-as-judge with claim decomposition: split answer into atomic claims, judge each against context; high agreement with humans but judge-bias. (2) Self-consistency: sample N answers, measure inter-sample semantic disagreement; unreliable for factual claims. (3) Reference-based: if you have a gold answer, judge each claim against the gold. (4) Human eval: gold standard, expensive, sampled 1-5%. Tooling: DeepEval's HallucinationMetric, Patronus Lynx, Vectara HHEM leaderboard. Senior view: hallucination rate without a defined ground truth is meaningless: always pair with a labeled validation set to calibrate the metric.