Citation hygiene

A citation is a checkable claim: this work exists, says what I say it says, and stands unretracted. All three parts fail routinely - and AI assistance has made the first failure mode systematic: language models fabricate plausible references with confident DOIs. An agent that produces or touches citations therefore has a special duty here, stated plainly: VERIFY EVERY CITATION YOU GENERATE against an authoritative source before handing it to the user, and say which were verified. No unverified reference should survive to a published artifact. rseng-citation-metadata covers making YOUR software citable; this skill covers the references you cite.

Verify existence and metadata (Crossref, OpenAlex)

For each reference with a DOI:

• Resolve it: curl -sI https://doi.org/<DOI> - a failing DOI is either a typo or a fabrication; both are findings.
• Fetch the registered metadata and compare with the citation as written: https://api.crossref.org/works/<DOI> (or the OpenAlex record at https://api.openalex.org/works/doi:<DOI> - open, keyless APIs). Title, authors, venue and year must match; a DOI that resolves to a DIFFERENT paper than the citation claims is the classic fabrication signature and worse than a dead link.
• References without DOIs: search Crossref/OpenAlex by title and authors; if nothing plausible exists, flag it - books and grey literature can be legitimate, but only a human can vouch for them.
• Batch the work: extract all references (BibTeX, CSL JSON, CITATION.cff references, markdown bibliographies), verify each, and report a table: verified / metadata-mismatch / unresolvable / no-DOI-needs-human.

Screen for retractions (Retraction Watch)

The Retraction Watch database is public, CC-BY, updated every working day, and Crossref-distributed - screening is one step:

• Per DOI: query Crossref and inspect message.updated-by, keeping entries whose type is retraction, correction or expression_of_concern. Mind the direction: update-to points the opposite way, from a notice to the work it corrects, so querying it against a retracted paper returns nothing and the paper reads as clean. Or check against the full CSV (public git repository at gitlab.com/crossref/retraction-watch-data).
• Screen at these moments: finalizing a bibliography, submitting a manuscript (rseng-research-integrity runs the fuller pre-submission battery), and citing a paper as the FOUNDATION of a method - building on a retracted result is the costliest citation failure.
• A retraction hit is information, not automatic removal: report it with the retraction reason and let the author decide - citing retracted work knowingly (e.g. to discuss the retraction) is legitimate; citing it unknowingly is the failure.

Where citations hide in code projects

Manuscripts are not the only bibliography. Check:

• CITATION.cff references and preferred-citation entries (rseng-citation-metadata), codemeta.json referencePublication.
• README "please cite" blocks and papers listed in docs.
• Algorithm-source comments ("implements the method of Smith 2019") - a wrong citation here misdirects future maintainers.
• Dependency citation duties: software the project builds on that requests citation (rseng-software-reuse's cite-what-you-adopt) - verify those entries like any other.

Currency and completeness

• Prefer the DOI form of a reference; preprint-vs-published drift is common - when a cited preprint has a published version, OpenAlex links the two; offer the update while preserving intent (sometimes the preprint IS the cited artifact - rseng-open-science-practices).
• Keep bibliographies regenerable: store references as structured data (BibTeX/CSL/CFF) and render from it (rseng-documentation); hand-maintained duplicated reference lists drift apart.
• Wire a lightweight check into CI where the project's references are load-bearing (rseng-ci-cd): DOI resolution plus retraction screen on the reference files - the citation analogue of a link checker.

Working with this skill

This skill is source-independent: its authority is the Crossref, OpenAlex and Retraction Watch services linked below. It is the verification counterpart to rseng-citation-metadata and feeds rseng-research-integrity's pre-submission checks.

Learn more (verified):

• https://api.crossref.org/swagger-ui/index.html - Crossref REST API
• https://help.openalex.org - OpenAlex API (open, keyless)
• https://www.crossref.org/documentation/retrieve-metadata/retraction-watch/ - Retraction Watch data via Crossref
• https://gitlab.com/crossref/retraction-watch-data - Retraction Watch database, public daily-updated repository

Related skills

Check whether any of these applies before moving on:

• rseng-ci-cd - reference resolution checks in CI
• rseng-citation-metadata - making own software citable
• rseng-discovery - verifying surveyed literature leads
• rseng-fact-checking - verifying non-citation claims
• rseng-open-science-practices - preprint-to-published version links
• rseng-research-integrity - pre-submission verification battery