Reference Change Desk. No AI guessing: each entry is classified with the resolver evidence shown."> Reference Change Desk. No AI guessing: each entry is classified with the resolver evidence shown."> Reference Change Desk. No AI guessing: each entry is classified with the resolver evidence shown.">
Pith. sign in

Pith Tools · Free · T1 deterministic audit

Reference-integrity audit

Upload a bibliography and every reference is checked mechanically against Crossref, OpenAlex, arXiv, and the Pith corpus; every resolved DOI is then checked against the retraction record (Retraction Watch and Crossref notices). A citation can resolve perfectly and still point at a retracted paper. No AI guessing which citations are real: each entry gets a deterministic classification with the resolver evidence attached, on a page you can share.

freeno account
150refs per run
RW+CRretraction check
T1deterministic
or paste references

Free, no account. Up to 150 references per run (larger uploads keep the first 150 and show the rest as truncated), 20 MB cap. Rate limit: 6 audits per 10 minutes per address. Results are public at an unguessable link. A full bibliography usually finishes in a minute or two; the resolver and the retraction check both call external registries, so pace depends on them.

01 What the classifications mean

The audit is resolver-backed and deterministic: the same file gives the same verdicts. It claims only what the evidence supports, so "no match found" is reported as unresolved, not as fabrication.

Registry match The reference's DOI or arXiv ID resolves and the metadata agrees.
Fuzzy match No explicit identifier, but bibliographic search finds a confident match.
Metadata mismatch The identifier resolves, but to a different paper than the reference claims. The classic hallucinated-citation signature.
Malformed identifier A DOI-like string that is truncated or syntactically invalid.
Fabrication risk A scholarly-looking reference with no match anywhere the resolver can see.
Unresolved No match, but nothing suspicious either: books, reports, and URLs are often unindexed.
Parse uncertain Entry too short or fragmentary to validate mechanically.

02 Standing of the cited works

Existence and standing are different questions. After each reference is resolved, its DOI is checked against Retraction Watch and Crossref notice events, plus a shared retraction cache. arXiv citations are covered too: withdrawn or retracted preprints are recorded under their DataCite DOI alias, and the audit derives that alias for every arXiv id it sees. Citing a retracted paper is the citation error referees punish hardest, and it is invisible to a resolver check alone.

Retracted The cited work carries a retraction or withdrawal notice. This breaks a clean verdict.
Concern notice A published expression of concern exists. Advisory, not a retraction.
Correction issued A correction notice exists for the cited work. Usually harmless; shown for completeness.

The audit also flags duplicate references (the same work cited under two entry numbers), advisory only. Finished runs carry a copyable proof block, a live SVG badge for READMEs, and a re-check button that reads the current retraction record without re-uploading anything.

Cited arXiv papers that Pith has not yet reviewed are queued for priority review, so auditing your bibliography also grows the reviewed record. Want an alert if a work you cite is retracted next year? Put your reference list on a standing watch with the living bibliography.

03 For agents and scripts

No cookies, no CSRF: POST a bibliography and poll the JSON. Same rate limit as the form. 

curl -s -F "file=@refs.bib" https://pith.science/tools/reference-audit.json
# → {"id":"ra_…","status":"queued","json_url":…,"badge_url":…,"recheck_url":…}
curl -s https://pith.science/tools/reference-audit/RUN_ID.json   # poll until status is completed

The stored run keeps the extracted references, their sha256, the audit version, per-finding resolver evidence, and the standing source with its as-of date, so any agent can re-run the same audit later and diff verdicts against the frozen evidence.