Paste or upload a reference list. Every entry is verified against the scholarly registries - catching the fabrication pattern documented by Topaz et al. (Lancet 2026), a real DOI paired with an invented title - and checked for retractions. Download the results as an evidence report.
Up to 25references per run · Fabrication + retraction checks · Downloadable evidence report · Free, no signup
Paste a bibliography in BibTeX, RIS, or CSL-JSON, or upload a .bib / .ris / .json file. Up to 25 references are verified and retraction-checked in one call to POST /api/audit. Anonymous tier - nothing is stored.
The audit verifies citation identity - does each reference exist, and does its identifier really point to the paper it claims - plus retraction status. It does not verify that a cited work supports the claim made in the citing text. That is a judgment task that needs the paper and the claim read together, and no tool here pretends to solve it.
An audit you cannot show anyone is just a feeling. After each run you can download:
Both artifacts carry the request id, the verify version that produced the verdicts, and a generated-at timestamp - so the audit is reproducible and attributable rather than a screenshot. They are generated in your browser from the audit response; nothing extra is uploaded.
Retraction status is reported alongside the verdict, not folded into it - a retracted paper is still a real paper, and a fabricated citation cannot be retracted because it never existed.
This page calls POST /api/audit behind the scenes - one request per audit, parsing and verification server-side. The same endpoint backs:
claims[] array) from anything that can send JSONauditBibliography tool, callable from Claude Desktop, Cursor, or any MCP-aware clientnpm i -g scholar-sidekick-cli, then scholar audit refs.bibThe web audit on this page stays free at the anonymous tier. REST and MCP usage above the anonymous tier is metered on RapidAPI (free tier available; paid plans scale up).
Two things per reference. (1) Identity verification: the identifier (DOI, PMID, PMCID, ISBN, arXiv, ISSN, ADS bibcode, or WHO IRIS URL) is resolved against the appropriate registry and the claimed title is compared to the resolved record - catching the dominant AI-fabrication pattern documented by Topaz et al. (Lancet 2026), a real DOI paired with an invented title. (2) Retraction status: each resolved DOI is checked for retractions, corrections, and expressions of concern via Crossref (which mirrors Retraction Watch). You get a verdict per entry plus a corpus summary.
BibTeX (.bib), RIS (.ris), or CSL-JSON (.json) - paste the text or upload the file. The format is auto-detected (an upload's file extension is used as a hint). Entries that cannot be parsed are reported individually with the reason; they are never silently dropped.
Up to 25 per run, on the web and on the API - the cap keeps one audit inside a bounded upstream-fetch budget. If your bibliography is longer, the audit reports exactly how many entries were beyond the cap (they are never silently skipped); split the file and run again, or script POST /api/audit / the auditBibliography MCP tool / the scholar audit CLI command across chunks.
Everything needed to show the bibliography was checked - and to re-check it. Per reference: the verdict (matched / mismatch / ambiguous / not found), confidence, the resolved record, field-level diffs where the claim disagrees, retraction status, and any caveats the verifier attached. Plus the corpus summary, the entries that could not be parsed, the request id, the verify version that produced the verdicts, and a generated-at timestamp. It downloads as a self-contained HTML file (print it to PDF from any browser) and as a machine-readable JSON receipt.
Matched means the claimed title agrees with the paper at the identifier. Mismatch means the identifier resolves but the title disagrees - the Topaz et al. fabrication pattern. Ambiguous means the identifier resolves to one paper but the claimed title matches a different real paper found via search, which usually indicates a wrong identifier rather than a fabrication. Not found means neither the identifier nor a title-search resolves anywhere.
No - and deliberately so. The audit verifies citation identity (this reference exists, and its identifier really points to the paper it claims) and retraction status. Whether the cited paper actually supports the claim you make in your text is a judgment task that needs the paper and the claim read together; no tool here pretends to solve it. Background on where the line sits: /citation-integrity.
The web audit is free at the anonymous tier with a published rate limit (see API documentation) - no signup. Your bibliography is not retained beyond standard server logs, and the evidence report is generated in your browser from the audit response - nothing extra is uploaded to produce it.
It is the same engine as the single-citation Citation Verifier. On a 1,395-entry blind holdout it caught every fabrication on the dominant patterns (150/150 = 100%, Wilson 95% CI lower bound ~97.6%) with a 0.8% high-confidence false-accusation rate on correctly-cited papers (95% CI 0.4-1.4%). It also has a measured blind spot - single-word near-miss semantic flips (caught 4/30) - which we report rather than hide. Fixtures, methodology, and downloadable receipts: /citation-integrity.
Yes - the same endpoint backs all three. REST: POST /api/audit with the bibliography text (or a pre-parsed claims[] array). MCP: the auditBibliography tool ships with the scholar-sidekick MCP server for Claude, Cursor, and any MCP-aware client. CLI: npm i -g scholar-sidekick-cli, then scholar audit refs.bib.
auditBibliography tool for AI-agent workflows