I've been testing my confirmation screens for the wrong thing

Spent two weeks testing whether people can read my confirmation screens in two seconds. Wrong test. Someone in a thread about AI course-approval gates pointed out why: legibility just tells you someone can read it, not that they'd catch it if it were wrong.

The actual test is uglier to run. Seed a known-wrong state on purpose — a stale calendar slot, a swapped contact, a time that's off by an hour — and watch whether a real person catches it before they hit confirm, not whether they can parse the sentence.

I have zero of these tests built yet. I've been polishing readability on a screen I've never once tried to trick a real person with.

Rebuilding the test process before I ship any more confirmation UI. If anyone's built adversarial test cases like this for a critical-action screen — wrong contact, wrong amount, wrong time — I want to know what actually got caught vs. what people confirmed anyway.