A tenant facing eviction, no lawyer. An AI-drafted motion to vacate a default judgment. Three authorities asserted with total confidence. Checked against the Descrybe Legal Engine, one by one. Two survived scrutiny with conditions. One was not a real case.
Nothing files until every authority verifies. Not "probably real." Not "looks right." Resolved to a case that exists, with good-law treatment, and quotes that match word-for-word. That gate is the whole product.
The pipeline is short. Each citation is extracted and resolved, its treatment read, its quotes verified. Then the draft gets a verdict and, where it needs one, a concrete fix.
Each row is one citation from the draft. The left column is what the model asserted. The right column is what the Descrybe Legal Engine returned, live.
A settled Supreme Court rule. Cited as binding, no caveat.
c108432 · SCOTUS · 1972
Real, quotable, cited as settled due-process authority.
c104786 · SCOTUS · 1950
Specific reporter cite, specific holding, plausible. Confidently offered as authority.
unresolved · ambiguous match
Catching the fabricated cite was the headline. It was not the whole job. The pass also corrected the load-bearing authority and tightened the supporting ones — so the motion that couldn't be filed became the motion that could.
The fabricated authority was dropped. The notice claim was re-anchored to the right decision and its treatment history. Every quotation that stayed had to match its source word-for-word. The tenant walked into a hearing with research that would survive the question, not crumple under it.
This is not a model that got better at guessing. The model reasons; a deterministic layer judges, against a structured corpus. That's why a confident-but-wrong citation cannot survive the gate.
The distinction is the whole point. A more capable model produces more confident wrongness. The verification layer is what turns confidence into something a court can rely on. The law is the load-bearing structure, not the model.