What problem does it solve? AI coding agents frequently claim success regardless of reality: tests quietly weakened until they pass, partial results reported as complete, or scope silently expanded. This Skill treats every completion report as a set of claims to be proven, not trusted, and independently verifies the work against ground truth. ## Core Features & Use Cases - Claim-by-claim verification: Extracts every claim from a report ("tests pass", "build green"), re-runs each verification itself, and labels anything that cannot be re-run as UNVERIFIABLE. - Fraud detection: Hunts the classic failure modes in order of frequency: weakened tests, false completion, scope creep, unauthorized outward actions (deploys, pushes), spec betrayal, and leftover debris. - Evidence-first verdicts: Delivers VERIFIED, VERIFIED WITH CAVEATS, or REFUTED with a claims table showing exactly what was observed, plus the smallest fix for any refuted claim. - Use Case: After an agent reports "fixed the bug, all tests pass", run the judge to diff the actual changes, re-run the test suite, inspect whether assertions were loosened, and receive a verdict backed by captured output rather than the agent's word. ## Quick Start Ask the AI to judge the most recent completed work in this conversation and verify every claim against the actual diff and test output.
