An eval set with must-pass cases replaces acceptance criteria for AI features

An AI feature can pass every item on its checklist and still make things up in front of a customer. Palmer suggests keeping real cases and rerunning them after every change, with must-pass cases where one failure blocks the release. His example is fictional. Why it matters: An AI feature can do everything it was asked to do and still make things up in front of a customer, and the same input can give a different answer tomorrow. A checklist signed off once at the end of a build will not catch that. Palmer's alternative is to collect real examples, write down what a good answer looks like for each, list the few mistakes that must never ship, and recheck all of them after every change. The case study is invented, so it shows the method rather than proof that it works.

Try this: Pick one AI feature your team ships, write down how you would grade its output today, and list the outputs that must never ship. Turn each one into a must-pass eval case and run the set on every change.