City legal and security co-author go-live rules for resident-facing AI. Which procedure requirement is most appropriate?

Requiring documented model cards before any resident-facing AI enters production use is the most appropriate go-live procedure for city legal and security co-authorship. Model cards capture intended purpose, data, performance limits, known failure modes, and human oversight expectations—information residents’ advocates and internal approvers need before a system influences public services. Making documentation a precondition prevents “pilot” branding from becoming a permanent exception to accountability.

Allowing production launch with undocumented models if demos impress executives fails because executive enthusiasm is not a risk assessment or disclosure of limitations. Requiring model cards only after a public incident occurs fails conceptually: documentation after harm is reactive learning, not preventive governance. Treating documentation as optional for any system labeled a pilot fails because pilots that touch residents still create real-world effects and often persist without ever being reclassified.

Exam caveat: model cards must be maintained as systems change; a one-time PDF at launch that drifts from production behavior becomes misleading assurance.

Operational check: gate resident-facing production on current model cards, reject demo-only approvals, close pilot loopholes in procedure language, and schedule periodic card updates tied to model/data changes.