A marketplace uses Claude to assign a condition label (new, refurbished, used, for parts) to listings. The prompt already states explicit criteria and the output format is consistent. On the eval set, nearly all errors are refurbished items labeled as used, and vice versa, often where sellers use informal wording. What is the most appropriate next step?

Answer: A — Add two to four boundary examples with varied wording and a short reason each, then compare per-category results

(a) is right: the criteria are already explicit and the remaining errors are a narrow boundary-judgment problem, which a few targeted, varied examples with reasons address cheaply; per-category results confirm the fix without hiding regressions. (b) adds tokens and latency to every listing for a problem that isn't multi-step reasoning. (c) inflates cost and risks teaching surface features. (d) buys capability everywhere for a narrow, fixable failure.