Question 1

Does my manuscript leave my browser, and what is stored server-side?

Accepted Answer

Not to the app's own infrastructure: there is no database and no application server, and the manuscript is read in your browser. TypeSafe rejects browser requests, so those calls go through a stateless host rewrite that stores and logs nothing. Citing sentences and cited-paper text do reach Anthropic and TypeSafe under your own keys, and the live site keeps only day, event, and referrer with no IP, cookie, or visitor id.

Question 2

Will it invent a quote and then tell me the paper supports the claim?

Accepted Answer

Not without proof, and not without you. Claude proposes a verdict and a verbatim passage, the app proves that passage exists in the fetched text, and a quote it cannot find is discarded and the citation flagged rather than shown as supported. Numbers in the citing sentence are checked against the paper in code. Jev scores supports, contradicts, or says nothing, and model output never overwrites a human verdict.

Question 3

What happens with paywalled or scanned sources — will a missing citation read as unsupported?

Accepted Answer

No, and that distinction is deliberate. A paywalled paper is checked against the abstract only, and an abstract silent on the claim is reported as needs full text, never as unsupported; you can attach a PDF to a row and every citing sentence is re-checked against the full text. Scanned PDFs are not read at all, and non-paper references are listed but cannot be checked automatically.

Question 4

Are those probabilities good enough to lean on, or do I need to read every row?

Accepted Answer

Treat them as triage. The README calls model verdicts a triage aid rather than a finding and says the thresholds have not yet been validated on a labeled set of biomedical citations. Either model works alone; with both, Jev scores the quote Claude already proved. Low P(supports) means Jev is itself unsure, which is a different signal from a contradiction.

Question 5

How are references matched, and what does it read?

Accepted Answer

Code reads the .docx or PDF, splits off the reference list and matches citation markers such as [12], superscripts, (12), ranges and author-year to sentences. References are matched by DOI, PMID, arXiv id or title in PubMed first, then Crossref and OpenAlex, so preprints, conference papers and non-biomedical journals are covered.

Question 6

What happens to my keys and usage data?

Accepted Answer

It is bring-your-own-key: citing sentences and cited-paper text go to Anthropic and TypeSafe under your own keys, kept in browser storage only and cleared when the tab closes unless you tick remember. The live site keeps just day, event (page opened, manuscript read, check started) and referrer, with no IP, cookie, visitor id or manuscript content.

Question 7

Which references can it not check?

Accepted Answer

References that are not papers, such as guidelines, regulations, websites or software, are listed but not checked automatically. A paper with no abstract on record whose publisher blocks browser downloads is marked so rather than guessed, and scanned PDFs are not read because the reader uses text position rather than OCR.

Paper Trellis Citation Verifier

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community