A municipal AI team cannot show which model version answered a resident complaint last month. What should the auditor emphasize?
Select an answer to reveal the explanation.
Short Explanation
When a resident complains, you need the receipt — which model version spoke. Keep lifecycle traceability so last month's answer is reconstructable. Strengthening pre-deployment bias and safety records won't tell you which model answered a specific past complaint.
Full Explanation
Ensuring traceability of deployed model versions through the AI lifecycle so past resident interactions can be reconstructed works because accountability requires knowing which model, prompts, and configurations produced a given municipal answer. When a resident complains about last month’s chatbot response, investigators need version identifiers bound to the interaction timestamp, not pre-deployment bias and safety testing records used to approve new model iterations. Traceability supports remediation, bias investigation, legal holds, and controlled rollback. The correct emphasis is durable version lineage across the lifecycle, not expanding transcript collection for quality sampling in the current production release, which cannot identify which model answered a specific past resident inquiry. Expanding collection of chatbot conversation transcripts for quality sampling in the current production release fails because sampling current conversations does not reconstruct which model version produced a specific answer from last month. Centralizing monitoring of aggregate accuracy and latency metrics to compare release performance over successive rollout cycles fails because aggregate metrics describe overall release performance, not which model version handled an individual past resident interaction. Strengthening pre-deployment testing records for bias and safety checks to support approval of each new model iteration fails because pre-deployment approval records do not trace a historical resident interaction back to the exact model version that answered it. Exam caveat: logging model IDs without retaining the referenced artifacts (weights, prompts, retrieval corpus versions) still leaves incomplete reconstruction—trace pointers must resolve to retrievable objects within retention policy. Operational check: emit model/registry version IDs on every resident interaction record, retain them per municipal retention rules, and periodically test retrieval of a historical interaction’s exact model and prompt package.