One vague prompt, seven upgrades, all eight prompts in full. Four add information; three change the process — and only one of those failures is easy to spot.

A quality gate is sufficient evidence, not a green test run — plus five review rules for code no human wrote.