Which method is most scalable for evaluating prompt quality across hundreds of test cases?

解析

• Bis correct: Anthropic's own evaluation guidance recommends combining scalable, automated grading methods: code-based/rule-based checks (unit tests) for objective and structured outputs, plus LLM-as-judge (using a model like Claude to score against a rubric via Likert/binary/ordinal scales) for subjective criteria — explicitly prioritizing volume over hand-graded quality.This is exactly what B describes, making it the most scalable choice for hundreds of test cases.
• Amanual review of every output is ground truth but does not scale and is reserved only for high-stakes final validation.
• C(A/B testing with real users) measures production outcomes rather than offline eval quality and depends on live user traffic, so it is slow and not suited to scoring a fixed test set.
• D(self-scoring) suffers from self-preference/consistency bias and is unreliable at scale — for LLM grading you want an independent judge and a rubric, not the model rating its own output.Only B is right, and no other option is defensibly correct, so the single-answer key is unambiguous.