Evidence Guard

About this skill

evidence-guard brings regulated-industry evidence standards — the same rigor used to pass medical and scientific MLR and peer review — to any AI-generated output. Before your agent ships documentation, READMEs, PR descriptions, API references, changelogs, or technical claims, evidence-guard forces it to prove every assertion against a real, verifiable source. The problem: AI agents write confident, plausible text that quietly contains unsupported claims — wrong version numbers, deprecated API references, performance figures with no benchmark, and documentation that drifts away from the actual code. A single model reviewing its own work rarely catches these, because the same bias that produced the claim approves it. evidence-guard fixes this at the level of substance, not just process. It runs a structured Claims-QC pass: it extracts every factual, technical, and quantitative claim from the output, classifies each one, verifies it against the repo or a citable source, grades the strength of the evidence, and flags risk patterns like version mismatches and doc-vs-code drift. Nothing passes the verdict gate until every critical claim is traceable. The result is a short, audit-ready Verification Note you can drop into a PR or a docs review. Built by PubsProToolkit, applying the evidence discipline of medical and scientific publishing to everyday agent output.