What problem does it solve? Written research records, findings documents, and incident reports often contain claims that drift from their cited sources: quotes get paraphrased, aggregate numbers go stale, grades like MEASURED or INFERRED get misapplied, and absence claims ignore retention policies. This Skill audits a record adversarially so factual errors and overclaims are caught before the document ships. ## Core Features & Use Cases - Claim re-derivation: Extracts every finding's claim, grade, and citation, then verifies quotes verbatim at cited path:line locations and re-runs every MEASURED command. - Grade and honesty audit: Checks that MEASURED maps to a rerunnable command, INFERRED is hedged, UNVERIFIED is stated plainly, and open questions carry no unsupported parentheticals. - Call-site and negative-claim verification: Greps for callers to catch gated or dormant code paths, and rules out prune/retention explanations before accepting absence claims. - Use Case: Before approving a findings document that cites a SQLite state database and CLI help output, run this audit to produce a verdict of ACCURATE, CORRECTED, or OVERCLAIMED per finding with path:line evidence for every verdict. ## Quick Start Adversarially audit the research record at docs/work/findings.md for factual accuracy, citation truth, and grade correctness, and report a verdict per finding.
