What problem does it solve? Technical documentation generated by an LLM-based producer (documenting-erp) can contain hallucinated claims, misread source code, or structural omissions that the producer itself cannot detect. This Skill acts as an independent second voice that re-reads the deliverable and challenges every assertion against the original X.13 sources, the SQL schema, and the DIVA reference corpus before publication. ## Core Features & Use Cases - Four error categories: Detects E1 (unsourced narrative claims), E2 (narrative misaligned with its cited source), E3 (structural omissions of fields, indexes, or foreign keys), and E4 (contradictions with the DIVA reference documentation), each with deterministic severity rules. - Hybrid deterministic + LLM pipeline: Regex-based detectors and vendored parsers (.dhsd, .dhsf, X.13 context extraction) handle structural checks, while an LLM pass evaluates semantic alignment between narrative and cited code context. - Strict read-only review: Produces review.md, review.json, and review.stats.json reports with a severity-ranked verdict (publishable / corrections needed / CA4 regression) and a producer coverage ratio, without ever modifying the deliverable or the reference docs. - Use Case: After generating documentation for the DAV module (entities CLI, FOU, ART, SOC, VRP), run this Skill to catch a planted misreading such as "CLI.Bloque = client actif" contradicting the cited source line, and receive an editorial synthesis plus a machine-readable report. ## Quick Start Review the UC-200 deliverable for the DAV module against the X.13 sources and DIVA reference docs, then show me the severity-ranked report and verdict.
