A county emergency-management AI has accuracy tests but no chaos or failure-mode tests for model-service outages. What should the auditor emphasize?

Including failure-mode and resilience testing for model-service outages in the AI test plan emphasizes that mission-critical county emergency AI must be validated beyond accuracy under healthy conditions. The correct emphasis works because accuracy benchmarks do not reveal whether fallbacks, manual procedures, degraded modes, and alerting work when the model service is slow, partitioned, or down mid-incident. Chaos and outage scenarios produce evidence that emergency staff can continue operations without unsafe reliance on unavailable AI outputs.

Assigning outage resilience validation to data-center continuity exercises for infrastructure failover fails because it shifts the AI solution's degraded-path design and rehearsal to infrastructure SLAs alone, leaving untested how the emergency AI itself behaves when the model service is slow, partitioned, or down mid-incident. Documenting manual override workflows for dispatch staff during model-service disruptions treats paperwork as a substitute for chaos or failure-mode testing; nobody has verified that those documented overrides are actually practical when the model service is down mid-incident and surge conditions are stretching emergency staff thin. Expanding offline accuracy benchmarking using additional historical incident datasets fails because more historical scoring still does not exercise live service failure, timeout handling, or continuity of emergency operations.

Exam caveat: a single simulated outage in a quiet lab may understate concurrent stress (radio traffic, surge intake, multi-site failure). Operational check: require documented outage test cases with expected fallbacks, measured recovery steps, and after-action updates to the emergency AI runbook before relying on the model in live county incident response.