What evidence do you require before shipping AI-generated code?

AI can produce a plausible change quickly, but shipping still requires evidence. My current minimum is: observable acceptance criteria, an explicit architecture boundary, a focused test that fails before the change, review of the actual diff, and release evidence from the production surface.

I am launching Vibe Coding Engineering on August 4 PT as a field guide to that workflow, based on real C++/Qt, QML, React, SvelteKit, AWS, payment, and delivery work.

Which gate catches the most AI-assisted coding failures for you—and which gate do today’s tools still skip?