- A lawyer filed a brief citing a case that never existed. An AI invented it. In March 2026 the Sixth Circuit fined the firm $30,000. The uncomfortable part: no smarter model would have caught it before it reached the judge. 🧵
- Filing-ready is a governance property, not a model property. A court does not care how confident the model was. It cares whether the case exists and says what you claimed. So we built CiteGuard as the verification layer around the AI, not another drafter.
- The core move: catching a fabricated case is deterministic Python, never an LLM. A brief cites Halstead v. Ferngate Holdings, 823 U.S. 1199. We do not ask a model if it looks real. We query CourtListener. Zero results. U.S. Reports has no volume 823. It cannot exist.
- That gap is the whole point. An LLM ruling "this seems fake" is a probability. A live zero-result lookup against real case law is proof of absence. Fabrication detection is exact, not a confidence score: the case is in the record, or it provably is not.
- So we split the engine deliberately. The existence check and the policy gate are deterministic. The one LLM step is the support-check: does this real opinion actually support the proposition it is cited for? And that step is allowed to abstain instead of answer.
- Abstain is the load-bearing word. When the model cannot establish support from the opinion text, it does not bluff a verdict, it routes to a human with the contradicting context. Confidence is cheap. Provable abstention is the part worth paying for.
- The gate is a deterministic invariant: filing-ready only when zero fabricated and zero unresolved citations remain. There is a unit test literally named test_fabricated_never_passes_gate. 0 fabricated citations can ever clear the gate. Not usually. Ever.
- One click exports a signed JSON audit certificate: each cite to its verdict, evidence snippet, real CourtListener URL, model and version or "deterministic", a UTC timestamp, the gate result. That receipt is what ABA Formal Opinion 512 and 300+ standing orders now demand.
- For legal-tech and AI-governance folks: for the one part of an AI system that must never be wrong, do you trust a stronger model, or a deterministic check the model cannot override? Where does your line sit? #LegalTech #AIGovernance #LitigationTech
- We built this as a demo, not a deployed product: 20 cached cases, a mocked document connector. If your team is figuring out how to file AI-drafted briefs safely, we would like to compare notes. Walkthrough: https://veriprajna.com/demos/legal-ai-citation-verification
Published on X · July 15, 2026
On social media
See this post on its original platform
In our archive