hallucination-checklisted

# Hallucination Check ## Purpose Verify that an LLM response is grounded in the context it was given. For each atomic claim in the response, check whether the retrieved context actually supports that claim. Anything not supported is flagged as ungrounded. This is the standard RAG safety net: the model retrieves N documents, generates an answer, and the verifier confirms the answer doesn't invent details that weren't in the documents. The skill is not a fact-checker against the open web — it only checks grounding *against the provided context*. If the context is wrong, the response can be perfectly grounded and still factually incorrect; that's a retrieval-quality problem, not a grounding one. ## Scope - **In:** atomic-claim extraction from the response, support-checking each claim against the context, structured ungrounded-claim report. - **Out:** retrieval quality (was the right context retrieved?), factual correctness against ground truth (was the context itself right?), citation formatting (does the answer have inline `[1]` markers?). Those belong to retrieval-evaluator and fact-checker skills respectively. Designed for `rag-architect` (debugging RAG pipelines) and `eval-engineer` (gating release of RAG-backed features). ## When to use - During RAG pipeline development, on a sample of responses, to find systematic hallucination patterns. - As a per-response runtime check before showing answers to the user (high-stakes domains: medical, legal, finance