“you click it and it breaks or you have made some graceful failure You should be able to do that on 100 people or 10 people and well then we're not going to get enough uh predictive power But I'm kind of okay and I paint a door that you get like 70% certainty”

Paint the door before you build the room

Before writing real code for a feature, show 100 users a button or screen for it, let them click through to a graceful failure, then survey them. A fake door to 100 people gives ~70% confidence on intent signal — enough to justify the real build or kill it. Hulls explicitly calls this the alternative to statistically-burdened percent rollouts: get the learning cheap, then commit.