StareBrain's confirmation screens were tested for the wrong thing

StareBrain shows you exactly what it's about to do before it does it — send a text, book a slot, whatever — and asks you to confirm. For two weeks I've been testing that screen for one thing: can someone read it in two seconds.

Wrong test. A thread about AI course-approval gates made this obvious in hindsight: legibility only proves someone can read the sentence. It says nothing about whether they'd catch it if the sentence were wrong.

The test that actually matters is uglier. Seed a deliberately wrong state — a slot that's already filled, the wrong contact, a time off by an hour — and see if a real person catches it before they confirm. Not "can you parse this," but "would you have stopped it."

I have zero of these built. Every confirmation screen I've shipped so far has been optimized for readability against a state I already knew was correct, which tells me nothing about whether it protects anyone from a state that's wrong.

Rebuilding the whole test process before I touch confirmation UI again. This is the actual job of a confirm-before-execute layer — not "did they see it," but "would they have caught it" — and I haven't been testing for it until today.

Building StareBrain in public, one adversarial test case at a time. Waitlist: starebrain.vercel.app/waitlist