Question 940
At Keystone Telecom, a technical contact center is improving network fault support during a controlled sandbox pilot (SF-G07-01). The administrator observes that a user reports an inaccurate refund estimate and the team cannot see which facts the agent used. Any fix must ensure that production changes remain gated by approval. Which continuing control best prevents this issue from recurring?
Explanation
Correct Answer
The preventive control directly addresses repeat failures: Use representative trace reviews during model and instruction updates; require the check in the next release review at Keystone Telecom (SF-G07-01). It complements the corrective action, inspect interaction details and trace events for grounding and executed actions, and protects the explicit condition that production changes remain gated by approval.
Incorrect Answers
This safeguard, “Review knowledge quality trends after ingestion changes; review it as part of periodic access and quality reviews”, reduces a different source-defined risk: Agent resolution scores fall when a new knowledge source is connected. The correction must ensure that policy terms come from the current approved revision. That risk depends on: Agent quality changes can arise from retrieval behavior as well as prompt text. The recurring problem here is a user reports an inaccurate refund estimate and the team cannot see which facts the agent used. The correction must ensure that production changes remain gated by approval. The originally selected control “Use representative trace reviews during model and instruction updates; review it after configuration or data-source updates” is more specific because interaction traces connect reported output problems to actual execution evidence.
This safeguard, “Alert on repeated action errors and confirm fixes with retests; review it before expanding channel availability”, reduces a different source-defined risk: One custom action intermittently fails but average conversation ratings remain high. The correction must ensure that service changes follow approved warranty conditions. That risk depends on: Action telemetry supports targeted repair that broad satisfaction scores cannot. The recurring problem here is a user reports an inaccurate refund estimate and the team cannot see which facts the agent used. The correction must ensure that production changes remain gated by approval. The originally selected control “Use representative trace reviews during model and instruction updates; review it after configuration or data-source updates” is more specific because interaction traces connect reported output problems to actual execution evidence.
This safeguard, “Keep tracing and privacy retention practices reviewed; review it with a documented exception process”, reduces a different source-defined risk: Operators receive complaint tickets but cannot correlate them to failing agent sessions. The correction must ensure that student records never cross program authorization boundaries. That risk depends on: Observability provides evidence of actual runtime behavior rather than anecdotes. The recurring problem here is a user reports an inaccurate refund estimate and the team cannot see which facts the agent used. The correction must ensure that production changes remain gated by approval. The originally selected control “Use representative trace reviews during model and instruction updates; review it after configuration or data-source updates” is more specific because interaction traces connect reported output problems to actual execution evidence.