Statistically rigorous real-world evaluation for robot policies. Plan how many rollouts you need, run them randomized and blinded from a phone at the robot, let fieldtrial switch LeRobot or openpi policies itself, and get a report that only says "better" when a pre-registered test does.
