Launch readiness and rollout gates

This is the lesson that fixes the thing that made the experiment produce a biased estimate. In Lesson 5.3, an A/B test was designed for the v1-to-v2 change. The result was borderline: +2.9 percentage points with a confidence interval that included zero. The decision was "hold for more data." But that estimate was wrong. The system has a shared cache: when treatment users populate the cache with better schema definitions, control users benefit. The control group was contaminated. The real effect is larger. This lesson covers how to diagnose violations of the independence assumption (when one user's treatment affects another user's outcome) and choose the right experimental design when standard A/B breaks. The starting point is recalling the assumption made when designing the user-level randomization experiment in Lesson 5.3.