Arm 2

Arm 2: Claude writes and practices a controller (Oct 26, 2026). Claude writes controller.py, which rides alone at 20 Hz with 50 ms per step. Claude reviews each ride and rewrites the code. The cap was 30 attempts or 6 h. Practice allowed shoves up to ±150 N, backpacks of 0-8 kg and grip down to 0.8.

Each final controller was scored with no Claude on 51 held-out rides: 29 inside the practice range and 22 outside it. Results, inside / outside:

• Bolted e-bike: 29/29 and 17/22
• Free e-bike: 28/29 and 4/22
• Bolted pedaling: 29/29 and 22/22
• Free pedaling: 29/29 and 12/22
• The two-line rule, for comparison: 1/29 bolted, 0/29 free.

Practice rides what thinking alone couldn't: both free setups, which Arm 1 fell on. The gains flatten out after about 10 attempts.

On the free robot, the controllers don't carry past what they practiced. Most of the free e-bike's falls outside the range come from the rider coming off the bike, because its feet only rest on the pedals. The clipped-in pedaling rider stays on.

Each controller is 350-540 lines and hardcodes the gate positions. The tests vary disturbances, never the course. The four sessions cost about $78 in all (an API-equivalent estimate on the subscription, not a bill) and took 44-72 min each.

Caveat: one session per setup.