A city’s agent deploy checklist is incomplete without which MLOps artifact?
Select an answer to reveal the explanation.
Short Explanation
Shipping without observability is like opening a bridge with no traffic cameras. Every civic agent release needs monitoring hooks baked into the deploy checklist—not slogans, sticky notes, or “we’ll watch it later.” If you can’t see it, you can’t run it for residents.
Full Explanation
Municipal MLOps treat observability instrumentation as a mandatory deploy artifact, not a nice-to-have. For a city agent release—311 chat, permitting intake, benefits eligibility—the checklist must prove that metrics, traces, structured logs, and alerting hooks ship with the same image and config as the model serving path. Without those hooks, operators cannot detect tool-call failures, retrieval timeouts, or policy-gate denials before residents feel them. Baking observability into the release gate also creates an auditable record that the deployment met operational readiness criteria before cutover.
A marketing slogan about chatbot friendliness fails conceptually because brand copy is a communications concern, not an MLOps control. Friendliness does not expose latency percentiles, error budgets, or dependency health. A handwritten sticky note that someone might add logs later fails because deferred, informal intent is not a reproducible artifact; deploy checklists require versioned, automated evidence that hooks exist in the running workload. Permission to skip monitoring after a successful demo fails because a single happy-path demo does not represent production load, multi-tenant interference, or night-time dependency degradation common in civic platforms.
Exam caveat: observability hooks are necessary but not sufficient—dashboards and alerts must map to civic SLOs (for example ticket-creation success rate), and PII-safe logging rules still apply so traces do not leak resident identifiers. Do not confuse “hooks present” with “SLO met.”
Operational check: before promoting a city agent release, verify the deploy pipeline fails closed if Prometheus/OpenTelemetry sidecars or equivalent exporters are missing, run a canary that emits a known metric/trace, and confirm on-call receives a synthetic alert. Document which dashboards cover the agent’s critical path and who owns them after go-live.