Frequently Asked
Questions

StrictCite ships its parser to you rather than asking you to ship your manuscript to it. The engine is Python compiled to WebAssembly, downloaded once and run inside your browser tab. Segmentation, parsing and normalisation all happen there, on your machine, using your CPU. When a single reference is too tangled for that regular-grammar parser to read cleanly, it gets a second pass from GROBID's real trained citation model — the actual weights and gazetteer data, copied from a real GROBID installation, decoded by Wapiti, the same open-source CRF engine GROBID itself depends on, compiled here directly from its own C source. GROBID's own code is Java and cannot run in a browser, so the tokenizer and feature extraction that feed that model are ours: read from GROBID's real source files and ported to JavaScript line by line, then checked field-for-field against GROBID's own live output on real papers until it matched. Not a generative AI, and not trusted blind either: every field it fills still gets checked against the registries afterward, same as everything else on this page.

In practice: when you open a forty-page thesis, the file is read into your browser’s memory, the reference list is located and extracted, and the body — your argument, your results, your unpublished findings — is discarded before anything is sent anywhere. There is no upload step to opt out of, because there is no upload.

This is an architectural guarantee rather than a policy one, which matters because policies get quietly revised and architectures do not. There is no endpoint on our server that accepts a manuscript. Not a disabled one, not a rate-limited one — none. There is nothing for us to change our mind about later.

Two things, and it is worth being precise about both.

Registry lookups. Titles, authors and DOIs go from your browser straight to Crossref, OpenAlex, DataCite, PubMed, Open Library and the rest. They travel directly — our server is not in that path and never sees them. Those registries learn that somebody looked up a paper, which is what they exist for.

The reference list and the records that came back. These go to our adjudicator, which applies the rule set and returns the verdicts. What travels is a list of citations to already-published work plus public catalogue records — never a sentence you wrote.

The adjudicator keeps none of it. It runs no database write of any kind: the reference list exists for the length of one request and is gone when the response is sent. Your bibliography is stored only if you explicitly save an audit to cloud history, which is a Pro feature and a button you have to press.

Why adjudicate on a server at all? Because the rules are the product — they are the part worth copying, so they stay on our side. That is a commercial reason rather than a technical one, and we would rather say so plainly than dress it up as something else.

You do not have to take our word for any of this. Open your browser’s developer tools, switch to the Network tab and run an audit: every request is listed, with its full contents. The engine is downloadable at /api/engine.zip — readable source, not a minified bundle, and byte-for-byte the code your browser just ran.

Thirteen, queried directly from your browser. Four are the agencies that issue identifiers — Crossref, DataCite, doi.org and Open Library — and are asked first.

Generative AI tools (ChatGPT, Claude, Copilot) frequently invent fake references: they generate real-sounding author names, valid journal titles, and fabricated DOIs that do not exist.

StrictCite resolves every identifier against the registry that issued it. If a cited DOI returns a HTTP 404 Not Found in official publisher records, or if the title text fails fuzzy similarity scoring against the canonical record, it is flagged as a suspected hallucination.

Yes! StrictCite produces standardized, normalized native exports:

• — .bib: Cleans syntax, standardizes citation keys, adds missing DOI & ISSN fields for direct Overleaf upload.
• — .ris: 1-click import into Zotero, Mendeley, and EndNote.
• — .json: For modern web bibliographies and Quarto/Pandoc pipelines.
• — .bbl: Accepted and repaired field by field.

Nothing is applied on your behalf unless two or more independent registries agree. A single registry's answer is shown as a claim attributed to that registry, not a settled fact — you decide whether to accept it, the tool does not decide for you.

This is not a hypothetical caution. Crossref, a real and generally reliable registry, has carried five duplicate, fabricated records for one of the most-cited papers in deep learning — all five resolving live, all dated years after the real paper. A tool that trusted any single registry's word would have confirmed the fake alongside the real one. Requiring agreement across independent sources is what stops one bad record from being treated as ground truth. See Registry Coverage for the full breakdown.

Plain pasted text works. Copy a reference list straight out of a PDF, a Word document, or a web page and drop it in — no formatting required. Or skip the copy-paste entirely and upload the PDF itself: StrictCite finds the reference section and splits it on its own, on your machine.

Real pasted text is messier than a clean export: paragraph breaks get lost, references run together with no separator, page headers and figure captions land in the middle of a citation. The parser is built specifically around these shapes rather than assuming a tidy input, and it is deliberately conservative about what it accepts — text that does not clearly read as a real reference is left out rather than guessed at and sent to the resolver as if it were one.

It watches papers you add, not your whole history: after an audit, add that paper's reference list to your watchlist (up to three at a time). Every Monday, a job checks four things against it: retractions, withdrawals, expressions of concern, and corrections, all from Crossref; whether a GitHub or GitLab link in a reference still resolves, is archived, or has new commits; whether a reference's journal is still listed in the Directory of Open Access Journals; and whether a cited preprint has been formally published or revised since. You get an email only when something changes, naming what changed. Each kind of check writes its own, so a busy week can bring more than one.

It is opt-in per paper, and there is a toggle in your workspace to turn the email off without removing what you are watching. Nothing is checked beyond what is listed above — a dataset, a non-open-access journal's internal standards, code hosted anywhere other than GitHub or GitLab, none of that is covered, and the workspace says so rather than implying otherwise. Free and the Deadline Pass check what you paste at that moment, once — watching a paper going forward starts at Researcher Pro.