Draft, .docx, .pdf or .bib · or
Paste a draft (markdown or plain text), upload a .docx / .pdf, or paste / upload a BibTeX .bib. Four checks per citation: does it exist, does its metadata match (authors vs. title), is it retracted, and does the cited abstract support the claim. Text is extracted in your browser — nothing is uploaded.
Without a key, the keyless check above runs in your browser (abstract-only, smaller model); untick it to only check that sources exist. With an OpenRouter key, an LLM compares each abstract with its claim instead — the optional higher-accuracy path. The keyless check is a first-pass screen, not a substitute for reading the sources — see what it can and cannot do. Either way, partial or unclear cases downgrade to cited, never pretend. Lookups go to free public indexes (Crossref, DataCite/arXiv, OpenAlex, Europe PMC, DBLP; Semantic Scholar only as a last resort), paced and cached locally so you hit fewer rate limits.
A single identifier, checked against the registries directly. No server, no signup.
Queries CrossRef / DataCite / arXiv live. A result of cited means the source exists — it does not mean the source supports your claim.
Measured, not promised. Every number is generated from published result files, misses and false alarms included.
Our labelled set: 34 samples — web app, no key, keyless check on (run 2026-10-21).
| Category | n | Caught | Misses | False alarms |
|---|---|---|---|---|
| Real citations (should pass) | 8 | 8/8 recognised | — | 0 |
| Fabricated | 8 | 8/8 | 0 | — |
| Chimeric (real title, wrong authors) | 6 | 6/6 | 0 | — |
| Retracted | 6 | 6/6 | 0 | — |
| Specific not in abstract* | 6 | 6/6 | 0 | — |
Public third-party datasets — existence / metadata level, same web pipeline, run 2026-10-21. Recall = problem references flagged; false alarms = real references flagged (not-checked items — rate-limited or unreachable — left out).
| Dataset (licence) | n | Recall | False alarms | Not checked |
|---|---|---|---|---|
| Badalova & Mayr 2026 — the July 2026 study's data (CC BY 4.0) | 104 | 22/33 | 15/71 | 0 |
| CiteMe benchmark v1, sample (CC BY 4.0) | 50 (+100 corrupted-field, in full results) | 22/25 | 0/25 | 0 |
| CiteTracer synthetic set, sample (MIT) | 135 | 42/90 | 0/45 | 0 |
| ICLR 2026 desk-rejected references, sample (MIT) | 50 | 39/50 | — | 0 |
On the July 2026 study's own data, veripaper raised fewer false alarms than any of the 5 tools tested there (21% vs. 28–75%) but caught fewer problems than 4 of them (recall 67% vs. 55–97%); it does not check venue, pages or partly wrong author lists. Web pages, software and standards have no scholarly record, so they count as false alarms here (UNVERIFIED).
The sample is small (6–8 per category), so read these as signals, not rates. The set is first-party (built and scored by us), registries change, and results vary with rate limits. *These six were built with the same deterministic check that flags them, so that row is not independent evidence of claim checking.
Full comparison, method and per-sample results → (also vs. RefExists, citegate and refchecker)
Five labels, three tags. No sixth option, no hedging.
The governing rule is downgrade-only: uncertainty always resolves toward the weaker label. A tool that cannot verify must say so.
veripaper is an open-source citation checker for research drafts. It looks up whether each cited source exists (CrossRef, DataCite, arXiv, OpenAlex, Europe PMC, DBLP, Semantic Scholar), resolves reference-list markers to identifiers, compares the authors you cited with the record, flags retractions, and screens whether the cited abstract supports the claim — with a small in-browser model by default, or an LLM with your own API key. Every citation receives an honest label from the contract above.
It never generates content and never pretends. When evidence is uncertain, it downgrades the label and leaves the judgment to the author.
Three horizons. Only Now is built; Next and Later are plans, not features.
veripaper will not estimate acceptance chances, simulate reviewers, or ghost-write rebuttals: those are predictions or generation the tool cannot verify.
Small helpers built along the way. They are not benchmarked; treat their output as a starting point only.
Can this become a paper? Live publication heat, venue map, and key papers — then a scored assessment. Numbers are free; judgment needs your key.
Heat counts come from OpenAlex (live). A thin count can mean a blue ocean or a dead end — the tool shows you the numbers; the field decides.
Pick the paper type and target venue. Get the checklist: where data is needed, how to collect it, and the bar for "done".
Describe the topic. Get a ranked venue shortlist with fit reasons — a starting filter, not a prediction.
Paste the reviews. Each point is split out, classified, and given a response template. Track what is addressed — export when done.
State is saved in this browser (localStorage). The tool structures the work; the science in each response is yours.
Where the draft went, what state it is in, and when the clock runs out. Per project.