A city wants each resident-facing foundation-model application to carry clear documentation of purpose, data, and known limits for governance reviews. Which artifact best satisfies that need?

Publishing programmatic model cards for each resident-facing foundation-model application best satisfies the city’s need for clear documentation of purpose, data, and known limits during governance reviews. A model card (or equivalent system card) records intended use, training/retrieval data categories, evaluation summaries, fairness and safety limitations, and operational constraints in a versioned artifact reviewers can cite. For municipal bots—311, permitting, benefits triage—that documentation is the durable evidence package governance boards expect when approving or re-certifying resident-facing FM apps.

A single informal chat message to the mayor with no version history is ephemeral, incomplete, and not auditable as a governance artifact. Undocumented prompt changes deployed directly to production nightly erase the link between documented purpose/limits and what residents actually receive. Marketing slogans that claim the bot is always accurate on every ordinance contradict responsible disclosure of known limits and create false assurance for reviewers and the public.

Exam caveat: Responsible AI governance artifacts for FM applications center on model/system cards (purpose, data, limitations), not ad-hoc messaging or marketing copy. Exam items often contrast documented cards with undocumented production drift.

Operational check: require a current model card before each resident-facing FM go-live or major prompt/RAG change, store cards in a version-controlled repo tied to app IDs, include purpose/data/limits sections that match production configuration, and present the card pack in governance review meetings.