Question 150
At fictional Longford Community Finance, a Snowflake data and AI engineering team is conducting a controlled model-deployment decision under an approved model lifecycle and release-control process. The separately composed technical incident is: A retailer runs a text classification function to route reviews, but category descriptions overlap and a large fraction of messages could legitimately belong to two labels. Assume only these documented facts and use safe, authorized production-change procedures; exact current Snowflake feature syntax and region or account availability require current review. Which repeatable assurance or control is most relevant to preventing this situation?
Explanation
Writer-proposed answer: B (UNVERIFIED)
Preliminary individual option explanations — NOT expert validated
Correct Answer
Review category drift, maintain a human appeal queue and revalidate classifications after taxonomy or model changes. Text classification quality depends on precise class boundaries, realistic labeled examples and a documented abstention policy.
Incorrect Answers
This response concerns Unauthorized tenant content in retrieval, which is different from the stated failure. A retailer runs a text classification function to route reviews, but category descriptions overlap and a large fraction of messages could legitimately belong to two labels.
Fine-tuning drift and baseline comparison is not the cause of the reported failure: A retailer runs a text classification function to route reviews, but category descriptions overlap and a large fraction of messages could legitimately belong to two labels.
Custom inference model deployment parity alone does not establish the needed evidence: An independently labeled confusion matrix shows acceptable class precision and recall, with ambiguous cases escalated rather than misclassified silently.