A policy that almost works still needs a human takeover when the mug tilts or the drawer sticks. Contributors on @axisrobotics watch the first attempt, seize control inside the short correction window, and turn that save into the next training batch. The browser sim loads without a GPU and without a physical arm on the desk. Score favors clean paths, hard scenes and skill spread rather than raw click count. Alliance tasks add a second reward rail once the trajectory is verified on chain. Kitchen pick and place still dominates volume while workshop insert and bathroom arrange keep the dataset from going stale. Slots close when the data quota fills, so late runs simply miss that stage. Replay after a first messy try usually produces the smoother demo the model actually needs. Robot general intelligence compounds only when failures stay in the loop instead of getting discarded.