Deferred evidence about evaluation

Editorial note: Evaluation is being handled separately, so this source passage has not been placed in the operating model. Teresa Torres is introduced in discovery-delivery-practitioner-evidence.md. The passage comes from 7c21461:ai/product-model/in/discovery-deliery-separation.md. Its wording has been edited for grammar and flow. The edit is intended to preserve the original claims. The citation is unchanged.

Torres then built an evaluation system from real production traces. It used human annotations, specific failure categories, LLM judges checked against human labels, a development set, and a larger test set. Measured regressions guided release decisions. During beta, she continued to review every coaching email by hand. (Product Talk)

Editorial comment: This is evidence about evaluation practice. It may later support the separate evaluation section. On its own, it does not define discovery or delivery.