bib-check: Catch Fake Citations and Wrong DOIs

What it does

An invented reference looks exactly like a real one: plausible authors, a fitting title, a real journal. This skill finds them the reliable way, by asking the registries that index scholarly work. Point your agent at a LaTeX paper or a .bib file and it:

• Looks up every entry. DOIs are resolved (Crossref, then doi.org for DataCite and other agencies), arXiv and PubMed ids are checked, and everything else is searched by title in Crossref, OpenAlex, DBLP and arXiv.
• Names each problem precisely: references no registry lists, DOIs that are not registered, DOIs that belong to a different paper (with that paper's title), arXiv ids that belong to another paper, and real titles cited with other authors.
• Works on your paper, not just your .bib: it follows \input, checks only the keys you cite, and says where each problem entry is cited (main.tex:15).
• Takes a second look before any verdict: for every entry the registries could not confirm, your agent searches the exact title on the web and tells you whether it exists elsewhere, resembles a different real work, or has no trace at all.
• Writes a Markdown report with counts, problems by severity, every found entry linked to its registry record, and the registries asked.

It says "no registry lists it", never "fake", and it never edits your .bib unless you ask.

What a run looks like

In our demo a fresh agent with only this skill got a 15-reference LaTeX paper with five planted problems. In 82 seconds it reported all five: two papers no registry or web search could find (one carrying the arXiv id of the Adam paper, cited elsewhere in the same .bib), a real title with invented authors (with the real ones), a DOI copied from another entry and a DOI with two digits swapped (each with the correct DOI). For the invented dawn-chorus paper it named the closest real study. It also cleared a real book the registries were unsure about.

How well it works

Scored once on 492 held-out references it was never developed on (two independent sets; the invented ones come from published compilations of fabricated citations found in ACL 2026 and ICLR 2026 submissions): real references found 189 of 190, invented references flagged 108 of 108, wrong DOIs reported 60 of 60, retracted papers found 70 of 70, and all 168 flags right. Most invented items are computer science; details and intervals on the docs page.

What's included

• SKILL.md: inputs, how to run the check, what each result means, the web double-check, how to report and fix
• scripts/bib_check.py: the check, one Python file, standard library only, network calls documented at the top
• examples/sample-refs.bib and examples/sample-report.md: a seeded bibliography and the report a real run wrote
• references/research-folder.md: the shared research/ folder convention

Part of the Aurora Lab research pipeline

Free, and works entirely on its own. Before a submission it pairs well with claim-check (every result backed by its data) and journal-house-style (any venue's author guidelines).