Evidence Guard is a skill that applies regulated-industry evidence standards to AI-generated output. It targets a specific failure mode: AI agents produce confident, fluent text that can contain wrong version numbers, references to deprecated APIs, performance figures with no backing benchmark, and documentation that has drifted away from the actual codebase. A single model reviewing its own work tends to approve the same errors it introduced.
Evidence Guard addresses this with a structured Claims-QC pass. It extracts every factual, technical, and quantitative claim from a piece of output, classifies each claim by type, verifies it against the repository or a citable source, grades the strength of the supporting evidence, and flags specific risk patterns including version mismatches and doc-versus-code drift. No output passes the verdict gate until every critical claim is traceable to a real source.
The skill is designed to run before an agent ships documentation, READMEs, PR descriptions, API references, or changelogs. The end product is a compact, audit-ready Verification Note suitable for dropping into a pull request or a documentation review workflow. The approach draws on evidence disciplines from medical and scientific publishing, specifically the rigor used in MLR and peer review processes, and applies them to everyday developer-facing agent output. There are no environment variables required and no transport is specified.