Each AI surface you test has a playbook: a primary metric, a day-1 starter, and the pitfalls that make weak results look good. After this module you can pick the right checks for chat assistants, summarizers, and classifiers, and run a concrete recipe for multi-turn testing.