What problem does it solve? AI-generated wiki entries can contain fabricated claims, wrong figures, or statements unsupported by their cited sources. This Skill audits wiki notes against the raw materials they cite, flagging unsupported, contradictory, or drifted claims before they spread. ## Core Features & Use Cases - Three isolated reviewers: A Blind reviewer (entry text only), a Grounding reviewer (entry plus raw source files and graph edges), and an optional External reviewer (confirmatory plus adversarial web searches) run with deliberately separated context to avoid anchoring bias. - Four-bucket triage: Findings are classified as decision_needed, patch, defer, or dismiss, merged across reviewers, and written to entry frontmatter as verify_status and findings fields. - Drift and provenance gates: Entries with missing raw files are marked drift_detected, entries without provenance are skipped, and decision_needed findings halt the run for user confirmation. - Use Case: After AI generates new wiki pages from research PDFs, run the verification across all sources entries to catch a hallucinated benchmark figure (e.g. 97.3% vs the actual 86.4% in the paper) before sharing the wiki. ## Quick Start Ask the agent to run /lumi-verify --all to fact-check every sources entry in the wiki against its cited raw files.
