Catches the hallucinated case before the judge does.
Free, open-source AI cite verifier for legal briefs. Paste a brief. Every citation is checked against CourtListener, GovInfo, and Justia. Quoted passages are fuzzy-matched against source text. Propositions are semantically checked. Out comes a Brigandi Shield Report PDF — verdict table, full reasoning log, ABA Rule 1.1 attestation language, SHA-256 input hash — plus an editable standing-order AI-disclosure cover sheet.
- Live: hallucination-shield.dashable.dev (when deployed)
- License: MIT
- Stack: Next.js 16 · TypeScript strict · Tailwind 4 · Anthropic SDK · @react-pdf/renderer · CourtListener v4 · Cloudflare Pages
- Stateless: briefs are not stored. Verification runs against the submitted text and the report is returned. Nothing is retained server-side.
In Q1 2026 alone US courts imposed more than $145,000 in attorney sanctions over fabricated AI-generated citations. The April 2026 Oregon "Brigandi" sanction order was $110,000 — for a single brief containing 23 fabricated cases and 8 invented quotes. 300+ federal judges have AI-disclosure standing orders. California's Bar is finalizing a verify-every-output rule across Rules 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3.
Every lawyer using Claude / ChatGPT / Harvey / Mike / any LLM now needs a verification layer. This is the seatbelt.
- Parse. Regex extracts cases (
Smith v. Jones, 123 F.3d 456 (5th Cir. 2010)), statutes (42 U.S.C. § 1983), and regulations (29 C.F.R. § 1604.11) from the input text or uploaded PDF. - Look up. Each cite is queried against CourtListener v4 (cases), GovInfo (USC + CFR), and Justia (fallback). Full opinion / statute text is pulled when available.
- Match quotes. Every quoted string in the brief is fuzzy-matched against the source text.
- Semantic check. Claude evaluates whether the cited authority actually supports the proposition asserted in the surrounding sentence.
- Roll up. Each cite gets a verdict — green (verified), yellow (review), red (defect: fabricated, misquoted, or off-point).
- Render. A Brigandi Shield Report PDF + JSON + standing-order disclosure cover sheet are produced.
pnpm install
cp .env.example .env.local # then fill ANTHROPIC_API_KEY
pnpm dev # http://localhost:3000Without ANTHROPIC_API_KEY the existence check, quote match, and Bluebook normalization all still work. The semantic check degrades to unknown and yellow verdicts proliferate.
See DEPLOY.md. Targets Cloudflare Pages via @cloudflare/next-on-pages.
10 pages/month per anonymous user. IP-hashed with a salt. No login. No retention.
- Replace your independent verification under ABA Rule 1.1 / Rule 3.3 / state equivalents. It assists, it does not substitute.
- Provide legal advice.
- Cover every secondary source. Treatises and Restatements are parsed but require manual review.
Issues and PRs welcome. Vertical forks (immigration cite check, family law cite check, IP cite check) actively encouraged — file an issue first so we can link to your fork from this README.
- Damien Charlotin — public AI-hallucination case database.
- Free Law Project / CourtListener — the public-interest case-law backbone that makes this possible.
- Mike — Will Chen's open-source legal AI work that inspired the "ship in a day" approach.
MIT © 2026 Stephane Boghossian. See LICENSE.