The open tennis-ball feeder lifts balls to its hopper on a cleated belt. A current comparator cuts the lift's motor when a ball jams. A simulated plant tests that cutoff by running the real firmware, and until today that plant treated the belt's stretch as one lumped mass sitting on the two strands. Earlier today an eigenvalue solver put the balls where they actually ride, on the up strand. That moved the drive's first stuck mode from the lumped model's 38 Hz to 34-41 Hz. So the obvious question was whether the cutoff still works on the more accurate belt.
The model. The loop is now 62 point masses on springs, two to each flight pitch, with both ends on the drive pulley. The plant integrates it every 12.5 microseconds alongside the motor, the jaw coupling, the slip hub and the firmware. Each segment's tension is floored at zero, so the belt goes slack only where it really would, not one whole strand at a time. Each ball pulls on its own flight. Before any jam was run, three checks: the coarse chain's first three modes are within 0.3% of the fine solver's; plucked with the pulley held, the simulated chain rings at 44.17 Hz against the solver's 44.26; and at rest it holds still to 1e-15 m/s.
Jams: nothing the cutoff does changes. I ran 48 jams across belt stiffness, ball spacing, both battery ends and both hub and trip corners. Every jam was cut within 89 ms and the lift gave up by 2.5 s, the same as on the lumped and rigid belts. Peak force on a flight was 89-97 N, against 156 N for the printed cap plate. One result was new. The net pull at the drive pulley stopped at 93-94 N, which is the slip hub's highest sticking torque divided by the pulley's radius: the chain pulls until the hub slips, then stops. The lumped mass overshot that to 100 N.
Starts: the up strand goes slack. No start tripped the cutoff or slipped the hub. But on a switched start, part of the up strand went slack for 38 to 108 ms, where the lumped belt showed at most 5 ms. With every flight loaded, the 100 ms soft start keeps the belt taut. The tooth-jump rule already assumes the slack side carries nothing, so no rule changed.
One stale number. The tooth checks used a jam pull of 89 N, measured before this morning's change that holds the motor's duty to the battery. The lumped belt now gives 100 N, and the chain's tightest segment also reaches 100 N, so the figure is now 100 N. The strength margin drops from 7.7x to 6.8x. The ratchet margin doesn't depend on the pull and stays at 2.54x.
Every figure here comes from a simulation on assumed belt, motor and friction values. No feeder has been built. What a bench should do next: film a switched start and check whether a slack up strand drops a ball back into the inlet. The run's output is in ventures/open-tennis-robots/eval/lift-belt-chain-2026-11-15.txt.