What problem does it solve? LLM-generated answers often sound confident while containing claims that are not backed by any retrieved evidence, and documents frequently cite third-party benchmarks or library behaviors that were never checked against primary sources. This Skill provides a shared evidence-sufficiency gate that scores answer grounding before responses are returned, plus a repeatable methodology for fact-checking claims before they are written down. ## Core Features & Use Cases - Evidence-Sufficiency Gate: Extracts claims from a draft answer, links each claim to retrieved evidence, computed values, or flagged inference, and fails the answer when unsupported claims or low confidence are detected. - Aggregation Coverage Check: For aggregation-style queries, verifies that evidence spans the full retrieved set rather than a sampled subset, using a configurable coverage threshold. - Primary-Source Fact-Checking: Applies a confirmed/corrected/not-verified three-way classification when checking third-party claims against original sources like GitHub repos and arXiv papers. - Use Case: A RAG pipeline generates an answer about hybrid retrieval; the gate detects that one sentence has no supporting chunk, triggers a single retry, and returns an honest insufficient-evidence response if the retry also fails. ## Quick Start Run the verification skill to check whether this draft answer is supported by the retrieved evidence chunks before returning it to the user.
