What problem does it solve? Codebases often embed algorithms, architectural choices, or business rules that silently violate established norms, standards, or best practices, and AI-generated audits frequently cite fabricated or remembered sources. This Skill validates a repository against real-world knowledge from scholarly APIs and standards bodies, and enforces that every finding is tied to a citation actually fetched during the session. ## Core Features & Use Cases - Research-grounded claim verification: Extracts falsifiable claims across algorithm, architecture, and business-logic layers, routes each to the right source class (arxiv, PubMed, OpenAlex, Crossref, Semantic Scholar, NIST, IETF/RFC, OWASP, ISO, sector regulators), and verifies against fetched evidence. - Anti-fabrication enforcement: fetch_sources.py appends every retrieved URL/DOI to an evidence log, and report_lint.py rejects any citation flagged fetched that never appears in it; ungrounded claims are reported as UNCONFIRMED, never as violations. - Adversarial refutation gate: Candidate VIOLATION/DEVIATION findings must survive independent refuter passes across correctness, citation-applicability, and severity lenses, with severity-aware vote aggregation enforced by the linter. - Use Case: Ask whether your lending module's APR calculation follows Regulation Z or whether your password hashing follows NIST SP 800-63B, and receive a severity-graded Markdown report where every verdict carries a real, fetched citation. ## Quick Start Ask the agent to run a base-in-reality audit on this repository and report any algorithms, architectural choices, or business rules that violate established standards with cited sources.
