No brief should file a citation nobody checked.
CiteProof is a citation and quotation checker for legal briefs. Paste a brief and every citation is resolved against public court records, every quoted passage compared against the opinion it claims — row by row, with the evidence on the page. It exists for the moment a partner signs: the report documents what was checked and when, so reasonable inquiry under Rule 11 is something you can show, not just claim.
How it started
In June 2023, a federal judge in New York sanctioned lawyers who filed a brief citing cases that did not exist — opinions invented wholesale by a chatbot, complete with confident citations to reporter volumes that had never held them. Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023) became the landmark, but it was only the first: sanctions orders over fabricated citations kept arriving, year after year, as AI writing tools spread faster than any habit of checking them.
The failure was never that lawyers used a new tool. The failure was that nothing checked the output — six citations sailed from a chat window into a filed brief without a single lookup. And the obvious fix had a trap in it: you cannot reliably catch AI guesswork with more AI guesswork. A checker that hallucinates its own verdicts is the same risk wearing a uniform.
So CiteProof was built the hard way on purpose: deterministic rules over public court records, every finding traceable to evidence on the page, every limit stated instead of smoothed over. No model in the pipeline, no silent passes, no score that flatters. The bet is simple — that the lawyers who check their briefs deserve tooling as serious as the duty they check against, and that a profession built on citation can hold its machines to the same standard it holds itself.
The problem it answers
Courts have sanctioned lawyers for filing briefs with fabricated citations — Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023) is the landmark, and sanctions orders over AI-generated citations keep arriving most weeks. The failure mode is always the same: nobody verified the cites before filing. CiteProof is the gate before filing — a second pair of eyes that never gets tired and never guesses.
Each week we read the new sanctions orders and run one through the verifier, row by row, in Sanctions Watch — free to read, one email a week at most.
What it does
- Existence. Every cited case is resolved against public court records. A cite that points nowhere reads as needs attention, never as verified.
- Identity. A fabricated cite can land on a real reporter location occupied by a different case. The name, court and year your brief claims are compared against the record — both sides shown when they differ.
- Quotes (paid). Quoted passages compared word-for-word against the opinion, side by side, with pin cites, match confidence, and word-level diffs for dropped negations and changed numbers.
- U.S. Code sections. Federal statute sections confirmed against the GovInfo Code edition — existence only, never currency or interpretation.
- A record for the file. Every paid check produces a Verification Record: counts, findings, the SHA-256 of the exact checked text, and a scannable audit code.
What it refuses to do
- No good-law verdicts. Whether authority is still good law belongs to Shepard's and KeyCite. Every resolved cite carries citing-history searches as a starting point, never a verdict.
- No proposition support. Whether a case supports your argument is legal judgment. The checker verifies the cite exists and says what it says — nothing about what it proves.
- No facts, no regulations, no state statutes. Reported opinions and the U.S. Code are the corpus. Everything else is counted, never hidden, and never checked.
- No silent passes. Anything the checker could not verify reads as couldn't-check with the reason stated. A false clean — a bad item marked verified — is treated as the worst possible outcome, in the code and on every page.
How a check works
| 1. Extract | Citations and quoted passages are pulled from your pasted or uploaded brief. |
|---|---|
| 2. Resolve | Case cites go to CourtListener; federal statute cites go to GovInfo. Citation strings only — never your brief. |
| 3. Compare | Names, courts and years against the record; quoted wording against the opinion page. Deterministic rules, no language model anywhere in the pipeline. |
| 4. Report | Verified, needs attention, or couldn't-check per row, with a counts line proving what was read — then a Verification Record for paid checks. |
The honesty rules it runs on
- Measured, never promised. Project measurements, October 2026 runs: 0 false cleans on 17 bad holdout items (rule-of-three upper bound about 18%, reported as a bound, never a rate). Identity comparison judges about 37.5% of full citations where a usable party name exists. Read the caveats on the accuracy ledger.
- Free checks are never stored. Pasted brief text is processed in memory; only counts and timestamps remain. Paid findings are kept 90 days, then deleted automatically. See the Privacy Policy.
- No AI reads your brief. Extraction and lookup run on parsed text and direct queries against public records. No model sits in the pipeline, and no brief text trains anything.
- It is not legal advice. Attorney review is required, Rule 11 diligence is yours, and the Verification Record documents what was checked — it proves nothing about your review.
Why it looks the way it does
A tool opened before filing has to feel like the rooms where filings get signed: quiet, exact, unhurried. So CiteProof is set like a chambers desk, not a dashboard — warm paper tones, a brass rule under the header, serif headlines with generous air around them. Nothing glows, nothing pulses, nothing begs for a click.
The restraint is deliberate. Colour is never the only signal: every verdict arrives as words first — verified, needs attention, couldn't check — because a tired eye at midnight must read the finding, not decode it. Dense evidence gets room instead of compression: side-by-side passages, full counts lines, research links grouped by the cite they serve. What shares a screen was decided by what a reviewer must see together, never by what fills space elegantly.
It is professional in the old sense: the interface stays out of the way of the judgment it supports. No session replays, no advertising trackers, nothing watching a brief being checked — and the same pages print cleanly for the file, because the file is where briefs live. Distinguished, if it earns the word, only the way a well-set brief is distinguished: by saying exactly what it means, once.
Who runs it, and how to reach us
CiteProof is run by its operator — billing and refunds through Paddle as merchant of record. System state is public on status, every user-visible change lands on the changelog. Questions, billing and refunds: support (billing and refund threads reply within 2 business days). Privacy requests: [email protected]. Everything else starts at contact.