# Prompt / model evaluation - Prompt keywords ≠ instruction-following; an instruction and its negation can contain identical keywords. - Test delivery, routing, data handling, and privacy separately. Mocked replies verify plumbing, not real model/provider behavior. - Evaluate actual outputs against task intent and representative failure scenarios, including conflicting evidence and permissions where relevant. Criteria must not depend on prompt phrasing. - Record actual model, observed outputs, outcomes, failures, limits. - Use existing evaluation infrastructure. Keep costly or variable evaluations outside the default fast loop unless required. - Evaluation unavailable or disproportionate → review the change and disclose the gap. No keyword/regex/snapshot substitute for behavioral evidence. - Retain reusable scenarios; consolidate overlap, discard scaffolding without hiding failures.