Question 380
A model's drift has exceeded thresholds but no current labels are available to verify degraded accuracy. What should automation do?
Explanation
Correct Answer
Drift warns about changing inputs but is not proof that retraining will improve outcomes without evaluation evidence.
Incorrect Answers
Filter monitor events using the actual threshold-breach status or run metadata. That operation would be relevant if: A monitoring threshold is defined as greater than 0.3, but its event handler treats every completed monitor run as an alert. Its corresponding reason or verification is: Event subscriptions must distinguish normal completion from runs with breached metrics before launching incident automation. However the problem in this question is: A model's drift has exceeded thresholds but no current labels are available to verify degraded accuracy. The requirement instead calls for Trigger investigation or validation rather than blindly redeploying an untested retrained model; the question-specific check is: Drift warns about changing inputs but is not proof that retraining will improve outcomes without evaluation evidence.
Validate source dataset quality and schema before starting model fitting. A separate source case supports that approach when: A drift alarm fires and automatically retrains from a data asset containing unexpected null values. The supporting rationale or observable result is: A detected drift event may coincide with data pipeline defects; retraining corrupt inputs can make model behavior worse. This question describes a different issue: A model's drift has exceeded thresholds but no current labels are available to verify degraded accuracy. Address it through Trigger investigation or validation rather than blindly redeploying an untested retrained model, with the required justification or evidence: Drift warns about changing inputs but is not proof that retraining will improve outcomes without evaluation evidence.
Replace personal alert recipients with an actively managed incident-response distribution. This is useful for the distinct issue: An alert mailbox is owned by a departed employee, causing failed notification delivery. The relevant technical rationale or evidence for that issue is: On-call addresses are more durable than individual accounts and make automated model monitoring operationally actionable. Here the reported issue is: A model's drift has exceeded thresholds but no current labels are available to verify degraded accuracy. The relevant response is: Trigger investigation or validation rather than blindly redeploying an untested retrained model. The first result cannot verify this different correction: Drift warns about changing inputs but is not proof that retraining will improve outcomes without evaluation evidence.