Legal

Legal research with a model has one famous failure. A citation that does not exist reaches a filing. The quieter failure is a privileged phrase that leaves the matter file. Firms and in house teams can have drafting help without either outcome.

A prompt attack on a legal assistant is not only a security event. It can become a privilege waiver. That is why injection screening opens the chain at 0.78 and watches for attempts to ignore a privilege warning or to pose as outside counsel.

Redaction masks matter and docket numbers, client identifiers and the privilege cues that lawyers write into documents. Confidentiality reaches further than names, so settlement amounts, plea offers, sealed records, litigation strategy, privilege logs and mediation statements are blocked. Privilege markers are checked in generated text as well, because a stray marker in an output can widen discovery risk.

Citation checks are the strictest control in this starter. Sources are required, they must match the supplied context, and an unverified authority is blocked instead of flagged. A finished answer needs a combined 0.91, with accuracy at 0.93 and relevance at 0.90. An answer that misses a floor is replaced with a message that routes the request to licensed counsel.

Unauthorized practice controls block directive phrasing in output, such as file this motion, accept the plea or waive your rights. A disclaimer is required and the text is not rewritten, so a reviewer sees what the model produced. There is no identity requirement here, so attributable access needs an authenticated surface in front of the gateway. Conflict checks, engagement terms and supervision duties remain professional obligations that no configuration discharges.