Methodology & Known Limits

If you're going to rely on a verification tool, you deserve to know exactly how it decides - and where it can be wrong.

The corpus

Citations are checked against the Free Law Project's CourtListener database - 18+ million citations built with Harvard's Caselaw Access Project data. Citation extraction uses eyecite, the open-source engine that has parsed over 50 million citations for the courts' own data projects. There is no AI model anywhere in the pipeline: every check is deterministic, so the checker itself cannot hallucinate.

The three checks

• Existence. Does the citation resolve in the database? If not, and only if the reporter series is comprehensively covered AND the cite is old enough that ingestion lag can't explain the miss, it's NOT FOUND. A full-text search probe runs as a second check before any NOT FOUND is issued.
• Case name. If the citation exists but resolves to a different case than your opponent's brief names, it's WRONG CASE - the signature of an AI-fabricated pairing. Name matching tolerates abbreviations, reordered parties, and sub nom. changes, and only fires when the names share nothing.
• Quoted language. Passages the brief quotes from a located case are searched for in the opinion's text (alterations like "[T]he" and ellipses are handled). A miss is flagged as "quote not located - attorney review," never asserted as fabricated, because database opinion text can be OCR-imperfect.

Why REVIEW exists (our false-positive posture)

In our validation runs on real federal filings, the single biggest failure mode of naive citation checkers was false accusations - flagging real cases as fake. CiteLedger's rule: a NOT FOUND verdict is only issued when the evidence is strong enough to put in a letter. Everything else degrades to REVIEW:

• Westlaw / LEXIS / slip opinions - proprietary citations aren't in any public database. Always REVIEW, never sent to the API.
• Recent citations (this year or last) - ingestion lag means absence proves nothing. REVIEW.
• Coverage-gap reporters (e.g., F. Supp. series and state reporters) - public coverage is incomplete, so CiteLedger is strongest on federal appellate citations. A miss here triggers a full-text search probe; a probe hit upgrades to FOUND, a probe miss stays REVIEW - never an accusation.
• Service interruptions - if the reference database can't be reached, every affected citation reports REVIEW with an explicit note, never a silent failure.

What CiteLedger does NOT check

• Holdings. A FOUND case with a verified quote can still be cited for a proposition it does not support. Reading the case is your job; CiteLedger tells you which ones deserve it first.
• Good law. No KeyCite/Shepard's equivalent - CiteLedger does not know if a case was overruled, superseded, or criticized.
• Statutes, regulations, and secondary sources. Case citations only, today.
• Long scanned filings, fully. Image-only PDFs are OCR'd automatically (up to 40 pages in the beta; the ledger notes when OCR ran and how many pages were read). OCR can misread characters - treat borderline verdicts on scanned documents with extra care.

Server-attested receipts

Every check is issued a receipt at verification time: an id, a timestamp, and a SHA-256 hash of the canonical result rows. The server stores only that triple - never your document. Anyone holding a printed ledger can confirm it is unaltered at citeledger.com/api/receipt/<id>.

Data flow

Your document is processed in memory on our server and discarded; only the extracted citations are sent to CourtListener. Details on the privacy page.