A new AI system checks its own legal citations before handing over an answer.
Researchers built GANDR, a two-agent system for legal question answering. One model, called the Drafter, writes an answer in a structured legal-reasoning format. A second model, the Critic, gets the same view as a human verifier and checks each individual claim against the passage it cites, producing an audit trace after every round. The system also enforces a strict rule: a citation only counts if it resolves to a passage the retrieval system actually returned, not one that merely sounds plausible.
On a 185-question legal benchmark where all six tested systems shared the same underlying model and retrieval setup, GANDR reached 70.8% strict accuracy, beating the next-best system by 11.3 percentage points. Turning off the protocol-anchored commit rule, the part that locks in the Critic's verdict, dropped accuracy by 22.7 points, which suggests most of the gain comes from the verification step itself, not better writing. Checked against two law-trained annotators, the audit's flag for under-supported claims scored an F1 of 0.84.
The researchers are upfront about the limits: the system's more detailed four-way verdict labels only weakly matched what the human reviewers picked, and the code is listed as available upon request rather than published outright.