How it works — with no backend
The registries are the backend. CiteWarden runs no server, no AI model, and pays nothing per check — it queries the same free public databases that publish the records in the first place: Crossref (the DOI registration agency, ~160M records), PubMed (the US National Library of Medicine), eCFR (the official US regulations text), OpenAlex, and CourtListener (optional token). A DOI either resolves on Crossref (HTTP 200) or it does not (404). That is not an opinion — it is a fact returned by the authority that mints DOIs.
- Extract — 10 citation families recognized: DOIs, arXiv IDs, PMIDs, U.S. Code, CFR, federal reporter cites, case names, UK acts, EU regulations & directives.
- Verify live — DOIs resolve against Crossref then OpenAlex; arXiv IDs through their registered DOI mirror; PMIDs against PubMed; CFR parts against the official eCFR.
- Verify offline — a 464-case landmark database (sourced at build time) catches exact reporter-cite matches and name matches.
- Forensic layer — impossible volumes, reporters before their first publication year, decision years that contradict the volume era, and a documented fabrication watchlist.
- Score — a transparent integrity score with a per-citation evidence trail. No black box.
Why it can't hallucinate a verdict
There is no AI in the decision loop. Parsing is deterministic regex. Verification is an HTTP status code from the authoritative registry. Forensics is arithmetic (volume 700 of U.S. Reports does not exist — that is arithmetic, not opinion). Deterministic code cannot invent a case out of nothing the way a language model can.
Where it CAN go wrong — stated plainly: it can miss an unusual citation format (a gap, not a false claim); a registry outage downgrades to "unverifiable", never to "fake"; and the fabrication watchlist flags by court record — every red card shows its exact reason and a source link so a human arbitrates. Amber exists precisely because "I don't know" beats a confident wrong answer. The whole engine is covered by 22 automated tests plus a live self-audit panel (top right) that re-proves its accuracy in your browser on every load.