Paper Trellis Citation Verifier

A free, standalone tool from Paper Trellis.

Try it: https://verify.papertrellis.com

Does the paper a manuscript cites actually support the sentence that cites it?

Upload a manuscript. The tool pulls out the reference list, pairs every citing sentence with the reference it cites, fetches each cited paper, and checks the pair two ways. You then record your own verdict — accept, reject, or other, with notes — on every citation, and export the table.

Built for peer reviewers and authors. References are matched in PubMed first, then in OpenAlex — so arXiv and other preprints, conference papers, and non-biomedical journals are covered too.

How a citation is checked

1. Code reads the .docx or PDF, splits off the reference list(s), and matches citation markers ([12], superscripts, (12), ranges, author–year) to sentences.
2. Code matches each reference by DOI, PMID, arXiv id, or title (PubMed, then Crossref and OpenAlex), then fetches full text when it is open — PMC for biomedical papers, the arXiv PDF for preprints — otherwise the abstract. Every result says which one was checked — "not in the abstract" is a much weaker finding than "not in the paper".
3. Claude proposes a verdict and the verbatim passage that backs it. The app then proves that passage exists in the fetched text; a quote it cannot find is discarded and the citation is flagged, never shown as supported. Numbers in the citing sentence are checked against the paper in code.
4. Jev (TypeSafe's System One model) reads the sentence next to that passage and returns a probability for supports / contradicts / says nothing. P(supports) is the reliability mark; low confidence means Jev itself is unsure.
5. You decide. Model output never overwrites a human verdict.

Either model works on its own; with both, Jev scores the quote Claude proved.

Privacy

There is no database and no application server. The manuscript is read in your browser. TypeSafe's API does not accept browser requests, so those calls are relayed by a stateless rewrite on the host (netlify.toml; the Vite proxy in development) that stores and logs nothing. The live site keeps an anonymous usage count (netlify/functions/count.mts): the day, the event (page opened, manuscript read, check started) and the site the visitor came from. No IP address, cookie, visitor id or manuscript content. Citing sentences and cited-paper text go to Anthropic and TypeSafe under your own API keys, which are kept in browser storage only (cleared when the tab closes unless you tick "remember"). A review is saved as a JSON file you download and can re-open later.

Manuscripts under peer review are confidential. Check the journal's policy and each provider's data terms before using this on one.

Run it

npm install
cp .env.example .env.local   # set VITE_CONTACT_EMAIL (used for Unpaywall lookups)
npm run dev

npm test runs the unit tests; public/sample-manuscript.txt is a tiny manuscript with one deliberately wrong citation to try it on.

You need an Anthropic API key and/or a TypeSafe API key. An NCBI key is optional and speeds up PubMed lookups.

Limits

• PDFs are read from the position of the text on the page (columns, real superscripts, running headers), which works on typeset journal PDFs and ordinary manuscript PDFs with selectable text. Scanned PDFs are not read. Text copy-pasted out of a PDF viewer is not reliable — upload the file.
• References that are not papers (guidelines, regulations, websites, software) are listed but cannot be checked automatically. Neither can a paper with no abstract on record whose publisher blocks browser downloads — the row says so rather than guessing.
• Paywalled papers are checked against the abstract only, and an abstract that is silent on the claim is reported as needs full text, never as unsupported. Add the paper's PDF on its row and every sentence citing it is re-checked against the full text.
• Model verdicts are a triage aid, not a finding. Thresholds have not yet been validated on a labeled set of biomedical citations.

License

MIT