A court clerk pilot needs an AI summarizer limited to local procedural manuals, not open-web knowledge. Which requirement should audit verify is enforced?

A court clerk pilot for an AI summarizer that must stay limited to local procedural manuals requires grounding and knowledge-boundary controls that restrict answers to approved municipal procedural sources. Clerks rely on jurisdiction-specific filing rules, timelines, and forms; retrieval tuning that prioritizes state case-law databases for broader legal context can sound authoritative while being wrong for this bench. Assurance therefore verifies retrieval allow-lists, exclusion of state case-law databases from retrieval, citation to approved manuals, and refusal or escalation behavior when the question is outside the corpus.

Retrieval tuning that prioritizes state case-law databases for broader legal context in summaries directly violates the knowledge boundary and increases operational risk by introducing non-municipal legal material. Citation formatting checks that standardize pin cites and exhibit references in clerk summaries do not restrict answers to approved municipal procedural sources; they only format references and leave grounding unenforced. Latency optimization through semantic caching of frequent procedural queries for faster clerk response times addresses speed, not grounding, and could serve cached responses without verifying they come from the approved municipal manuals.

Exam caveat: grounding controls must be technical and procedural—prompt instructions alone are weak against model drift and tool misuse. Operational check: test out-of-corpus prompts for refusal, verify citations resolve to current manual versions, confirm change control when manuals update, and document who may expand the approved source set.