What problem does it solve? Academic papers often contain hallucinated references, wrong-context citations, author name errors, year mismatches, and phantom DOIs that survive manual review and can sink a submission. This Skill audits every \cite{...} in a LaTeX manuscript against real web sources (DBLP, arXiv, ACL Anthology, OpenReview) before submission. ## Core Features & Use Cases - Three-layer verification: checks each cited entry for existence at the claimed arXiv ID/DOI/venue, metadata correctness (authors, year, venue, title), and context appropriateness (whether the cited paper actually supports the surrounding claim). - Cross-model independent review: sends each entry to a fresh gpt-5.5 reviewer thread via Codex MCP with mandatory web lookups, producing per-entry KEEP / FIX / REPLACE / REMOVE verdicts. - Structured artifacts: always emits CITATION_AUDIT.md (human-readable report) and CITATION_AUDIT.json (machine-readable ledger with verdict, reason codes, and input hashes) for downstream verifier gates. - Optional modes: --uncited detects bib entries never cited (detect-only), and --soft-only freezes the bib file and proposes sentence rewrites in the .tex files instead of bib mutations. - Use Case: Before submitting a paper, run the audit to catch a citation of Self-Refine used to support a claim it actually contradicts, an anonymous-author placeholder entry, and an arXiv preprint that should now cite the NeurIPS version. ## Quick Start Run the citation audit on my paper directory to verify every reference in references.bib against real sources and generate the audit report.
