SELF-SUSTAINING DATA LOOP the wind tunnel, productized ① FLEET deployed robots act ② FOUNDRY auto-collect · label · filter quality ③ TRAIN foundation model ④ WIND TUNNEL real-robot eval gate · locate fail human teleop / labeling
35%

This is the org chart the wind tunnel produces. The four stages stop being separate teams with quarterly handoffs and fuse into one fast ring: a fleet acts → a data foundry auto-collects, labels, and quality-filters → training folds it into the foundation model → the wind tunnel gates it on real robots and locates the next failure → back to the fleet. The amber human-in-the-loop node is the fuel you're trying to burn off: crank automation and it shrinks while the ring spins faster. Past escape velocity it never fully disappears, but it stops being the bottleneck.