A support-answer assistant collects thumbs-up/down ratings from agents. A dashboard shows satisfaction declining for two months. The prompt team keeps adjusting the prompt and checking a handful of examples by hand. What is the main architectural gap?

Answer: A — Structured feedback: downvotes with a reason and correction, sampled into a versioned eval set run before each change

Feedback that ends in a dashboard improves nothing; (a) turns it into regression data so changes are tested against real failures. (b) iterates faster without measurement. (c) guesses at the cause. (d) collects more unstructured signal without closing the loop.