There is a quiet sleight of hand in most legal AI marketing. The product generates text with a language model, then “verifies” it — with another language model. A second opinion from the same species of system that invented the error is not verification. It is a coin flipped twice.
The sanctions record makes the distinction concrete. The briefs that drew fines this year did not fail for lack of confidence — the fabricated cases came with reporter numbers, pinpoint pages, and quotations. They failed because nothing deterministic ever stood between the model's output and the court. A model can be fluent about a case that does not exist. A database cannot.
Verification has to be a lookup
A legal citation is one of the most verifiable objects in any profession. 347 U.S. 483 either is Brown v. Board of Education or it is not. There is no probability distribution over that fact. So verification should not be a judgment call — it should be an exact match against an authoritative table of every reporter citation that exists.
That table did not exist in a form we could query at interactive speed on every proposed citation, so we built it. We ingested the Free Law Project's bulk citation data — the citation graph behind CourtListener — into our own index: 16,307,430 reporter citations, each mapping a volume, reporter, and page to a real, linkable opinion.
Every citation an agent proposes in Crebral Legal is resolved against that index before you accept it. Brown resolves. Miranda resolves. Roe, Celotex, the workhorse citations of everyday motion practice — they resolve to specific opinions with links you can open (the links in this sentence are exactly what our index returns for those cites). A hallucinated cite, however fluent, resolves to nothing, and the interface says so before the citation enters your draft, not after opposing counsel finds it.
Click-to-confirm, not trust-by-default
Verification also has to be visible to be worth anything. In our editor, Cite Authority marks every citation with its verification state. Proposed citations are auto-verified before you accept them; verified cites carry a badge that links to the underlying authority. The workflow is click to confirm, never trust by default — because the lawyer signs the filing, and the lawyer deserves to see the evidence, not a vibe.
When a tool cites a case, ask what checked it: a lookup you can audit, or the same model that wrote the sentence.
Independent models still matter — we run consensus across six providers because holdings, unlike citations, require judgment to characterize, and adversarial cross-examination between unrelated models catches mischaracterization. But the existence question — is this a real case — is never delegated to a model in our system. Models propose. The index disposes.
Accountability is an engineering decision
The profession spent this year learning that “the AI checked it” is not a defense a court will accept. We agree with the courts. Verification that can be audited — a named index, an exact match, a link to the authority — is the only kind that deserves the word. Everything else is generation wearing a costume.
Our citation index is built from Free Law Project / CourtListener bulk data, refreshed quarterly. The case links in this article (Brown, Roe, Celotex) are the opinions our production index resolves those citations to. Hallucination rates for incumbent tools are from Stanford RegLab (Magesh et al.).