In 2023, lawyers were sanctioned for filing a brief of cases ChatGPT invented. This agent is built so that cannot happen — not because the model promises, but because the system checks:
| Bounded corpus | Real published opinions from CourtListener in a Sanity dataset. The banner above is written by a census script from measured values — never typed by hand. |
| Mechanical quote check | Every claim must carry a verbatim quote. Code — not the model — verifies it exists letter-for-letter in both the text the model was shown and the stored opinion. Failures are rejected and stripped. |
| Withheld summaries | If zero claims survive verification, no summary is shown. An unsupported answer never wears this tool's credibility. |
| Server-truth labels | Case names, dates, courts, and links come from the database record — the model cannot invent a URL or mislabel a case. |
| Honest edges | "Not in corpus" means not in THIS collection, never that no authority exists. Gaps say what the corpus can't answer; truncation is disclosed per answer. |
Try it yourself: the verify tab's "infamous fake case" chip is Varghese v. China Southern Airlines — the actual fabricated case from the sanctions story. Every AI chatbot will summarize it for you. This one structurally can't.