Question 480
A team uses only a model-as-judge score to approve a critical regulated prompt. The proposed fix needs to work without changing unrelated approved resources. A controlled test should demonstrate that the release record includes critical manual judgments and machine checks. What should the release engineer do before proceeding?
Explanation
Correct Answer
A model-graded score is not universal proof of factual or policy correctness. A successful run shows that the release record includes critical manual judgments and machine checks.
Incorrect Answers
Uncontrolled grading variability undermines prompt comparison decisions. A model-graded score is not universal proof of factual or policy correctness.
Changed evidence retrieval can alter outputs independently of prompt text.
Semantically similar text may still be unusable for a structured API.