Build a 20-question evaluation set for a live-web feature

Define expected answer elements, freeze source evidence, and review a live-web feature with a small evaluation set that keeps failures and missing scores visible.