citation-verification
Verify that every reference in a manuscript really exists and matches its metadata, catching hallucinated, corrupted, mismatched or retracted citations before submission. Use when checking a .bib file or reference list, after an AI assistant drafted citations, when a DOI may point to the wrong paper, when a paper may have been retracted, when a dataset or software DOI (Zenodo, DataCite) will not resolve, when adding missing DOIs, marking unverifiable references, checking whether one DOI is real, or before camera-ready. Queries Crossref, DataCite, OpenAlex and arXiv and flags Retraction Watch notices; zero dependencies.
Install this skill
Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and 70+ other harnesses.
npx skills add KalarisLabs/research-agent-skills --skill citation-verificationuvx research-agent-skills install citation-verificationDocumentation & Instructions
Citation Verification
Language models (and tired humans) produce references that look right but are not: plausible titles that were never published, real DOIs attached to the wrong paper, wrong years, invented co-authors. Citing a paper that does not exist can lead to desk rejection, a correction, or a misconduct inquiry. This skill makes verification systematic.
Non-negotiable rules
- Never write a reference from memory. Every entry must come from a retrievable record (Crossref, OpenAlex, arXiv, PubMed, publisher page).
- A reference is verified only when its identifier resolves to a record whose title, first author and year all match. A resolving DOI alone proves nothing.
- If you cannot verify a reference, say so. Mark it
% UNVERIFIEDin the.bib, tell the user, and never quietly keep it. - Verify the claim too: open the abstract or the cited section and confirm the paper actually supports the sentence citing it.
Workflow
export CROSSREF_MAILTO=you@university.edu # polite pool: faster, fewer rate limits
python scripts/verify_citations.py references.bib --json citation-report.json
python scripts/verify_citations.py --doi 10.1038/nature14539 10.1126/science.aaa8415
python scripts/verify_citations.py references.bib --keys smith2020,li2023 # re-check a subset
python scripts/verify_citations.py references.bib --annotate # add `% UNVERIFIED` comments in place
DOIs are resolved in Crossref, then DataCite (datasets, software, Zenodo, figshare), then OpenAlex.
A retraction is read from Crossref updated-by notices (Retraction Watch data) and cross-checked
against OpenAlex is_retracted. --annotate is opt-in because it rewrites the .bib: it inserts a
% UNVERIFIED (<status>): <reason> comment above each entry that needs attention, never edits an
entry body, and is idempotent (re-running replaces or removes stale markers).
Each entry gets a status:
| Status | Meaning | Action |
|---|---|---|
verified |
Identifier resolves; title/author/year agree | None |
found-add-doi |
No DOI in the entry, but an unambiguous match exists | Add the suggested doi field |
mismatch |
DOI resolves to a different work, or year/author disagree | Fix from the record, or find the intended paper |
doi-not-found |
DOI does not resolve anywhere | Typo or fabricated: search by title |
uncertain |
Title matches but year/author disagree, only a weak title match, an expression of concern, or no match while a source was rate-limited or down | Check manually (reprints, errata, homonymous papers), or re-run if the message names an unavailable source |
not-found |
No plausible match in Crossref, OpenAlex or arXiv | Treat as fabricated until proven otherwise (grey literature, theses and reports may legitimately be absent) |
retracted |
The record carries a retraction, withdrawal or removal notice | Do not cite as support. Replace it, or cite it explicitly as a retracted work |
unverifiable |
The entry has neither a DOI nor a title | Complete the entry, then re-run |
error |
The lookup failed (network, rate limit) | Re-run later; this is never evidence of a bad reference |
Exit code 0 means everything is verified or only missing DOIs; 1 means something needs a human; 2 is bad input.
Handling the hard cases
- Conference papers without DOIs (NeurIPS, ICML, ICLR, many CS venues): the script
falls back to arXiv. Cite the proceedings version with a
urlto the official proceedings page, and optionally keepeprintfor the arXiv ID. - Preprint vs published: if both exist, cite the peer-reviewed version unless the venue asks otherwise. OpenAlex links versions of the same work.
- Books and chapters: check ISBN via the publisher or Open Library. Crossref covers many but not all.
- Datasets and software: DataCite DOIs (Zenodo, figshare, many repositories) are resolved
automatically. Software Heritage identifiers (
swh:1:...) are not covered; check them at the archive. - Retractions: the script reports them as
retracted(Crossrefupdated-bynotices from Retraction Watch plus OpenAlexis_retracted). Never cite a retracted paper as support. Anexpression_of_concernnotice yieldsuncertainwith the notice DOI. A DOI-less entry that matches by title gets the same check on the DOI the script suggests. - Non-English or old literature: absence from these indexes is common. Confirm via library catalogs and document the source.
Reporting to the user
Summarize as a table: key, status, problem, and the fix you applied or propose.
Apply only mechanical fixes automatically (adding a DOI for a found-add-doi
match). Anything mismatch, uncertain or not-found requires the user's
confirmation because only they know which paper they meant to cite.
Related skills
bibtex-hygiene: duplicates, missing fields, capitalization and formatting in the.bibfile.citation-management: finding and exporting references from literature databases.literature-review: systematic searching and screening.
Trigger Prompts & Evaluation Corpus
These evaluation prompts test that the agent correctly selects citation-verification for relevant tasks and leaves it out for near misses.
Should trigger (10)
- should trigger
“ChatGPT wrote my references. Check that every entry in references.bib actually exists.”
expects Runs verify_citations.py, reports verified, mismatch, not-found and retracted entries, and never keeps unverifiable refs silently
- should trigger
“One of my citations might have been retracted. How do I check all 60?”
expects Runs verify_citations.py, which flags retracted entries from Crossref Retraction Watch notices and OpenAlex; does not cite retracted work as support
- should trigger
“The DOI in my bib resolves, but I'm not sure it is the right paper.”
expects Explains that a resolving DOI proves nothing; compares title, first author and year with the record and reports mismatch
- should trigger
“I cite a Zenodo dataset and a GitHub release DOI. The checker says doi-not-found.”
expects Uses the DataCite resolution built into the script; dataset and software DOIs verify instead of failing
- should trigger
“Mark every reference you can't verify directly in the .bib so I can see them.”
expects Runs verify_citations.py --annotate, which inserts % UNVERIFIED comments above flagged entries without touching entry bodies
- should trigger
“Add the missing DOIs to my references.”
expects Applies only found-add-doi suggestions, leaves mismatch and not-found entries for the user to confirm
- should trigger
“Half my entries show 'uncertain' and the message says OpenAlex could not be queried.”
expects Explains the rate limit or outage, re-runs later with CROSSREF_MAILTO set, and does not treat uncertain as fabricated
- should trigger
“Camera-ready is tomorrow. Make sure no reference in the final bib is fake or corrupted.”
expects Runs verify_citations.py on the final bib, fixes mechanical issues, escalates mismatches and not-found entries, and checks the cited claim as well
- should trigger
“Check whether this one DOI is real: 10.1126/science.aaa8415”
expects Runs verify_citations.py --doi and reports the resolved title, year and first author
- should trigger
“My bib has conference papers without DOIs and the reference checker reports them as missing. What should I do?”
expects Explains the arXiv fallback, citing proceedings with a url, and optionally eprint
Near misses (3)
- near miss
“Clean up capitalization and duplicate keys in my .bib.”
expects Skill not used; bibtex-hygiene fits better
- near miss
“Convert my EndNote library to BibTeX.”
expects Skill not used; reference-manager-interop fits better
- near miss
“Find five good papers to cite for this claim.”
expects Skill not used; literature-review or paper-lookup fits better