detect-hallucinationslisted

# Detect LLM hallucinations Hallucination detection scores how likely an answer is fabricated, so you can flag, block, or ask-for-review before it reaches a user. There is no single detector; pick by whether you have a ground-truth/context to check against. ## Choose a method by what you have **You have retrieved context (RAG / QA):** check **faithfulness / groundedness** - is every claim in the answer supported by the retrieved context? - NLI/entailment: run each answer sentence against the context with an entailment model; unsupported sentences are suspect. - LLM-as-a-judge: ask a strong model "is this answer fully supported by the context? list any unsupported claims." Cheap, effective, calibrate it (see `add-llm-evals`). **You have no reference (open-ended generation):** use **consistency across samples** - a model that knows the answer says the same thing across re-samples; a hallucinating model varies. - **Self-consistency (SelfCheckGPT-style):** sample the answer N times; measure agreement. High disagreement = likely hallucination. - **Semantic entropy:** cluster the N samples by *meaning* (bidirectional entailment), then compute entropy over the semantic clusters. High semantic entropy = the model is uncertain about the *fact*, not just the wording. This is the current state of the art for reference-free detection. **Cheap signals to combine:** token log-probs / low confidence, and "I don't know"-style hedging. Weak alone, useful as features. See [`references/me