bib-audit
Bib Audit
Two jobs, in priority order. First, flag hallucinated references, authors and bib items, meaning works that don't exist, identifiers that name a different paper, and authors who aren't on the paper. Second, correct badly formatted references, by resolving every field from the registrar of record instead of hand-authoring it.
The core rule is to never author bib fields, only resolve them. Every reference should trace to a DOI or arXiv ID.
Why this matters more than tidiness. The whole idea is that authors run this before submitting, so that reviewers never have to spend their time spot-checking a bibliography to decide whether an LLM wrote the paper. Right now checking someone's references is unpaid cleanup for a mistake you didn't make, and reviewer attention is the scarcest thing in the process, so that's a lousy way to spend it. If everyone gates their own submission on this, the doubt goes away as a category. So the tool is built author-first: it's read-only, it exits non-zero for CI, and it tells you how to fix each finding rather than only naming it. Using it to audit someone else's submission works and is documented below, but that's the fallback, not the goal.
That priority sets the design bias too, which is that a false accusation is worse than a miss. A missed bad reference is one entry a reviewer might still catch, while a false "fabricated" verdict on a correct entry destroys trust in the whole report, and in a review it puts an integrity allegation next to someone's name. When in doubt the tool downgrades to advisory rather than escalating.
Both situations differ only in what you do with a finding:
- Your own paper (draft, pre-submission, camera-ready). Fix in place. Replace fields from the canonical source, add missing identifiers, re-run until the audit is clean, then wire it into CI as a gate.
- Someone else's paper (peer review, reading a preprint, checking a collaborator's draft). You can't fix it, so the output is evidence. Resolve each finding to "identifier names no paper" or "identifier names a different paper" before saying anything, quote the specific entry, and read the review-etiquette note below, because a citation-manager glitch and deliberate fabrication look identical in the audit output.
Quick start
Pick the entry point by what you have.