# Physical arm stop on a lost gate reply — 26 September 2026

Company-authored engineering evidence. One physical SO-101 elbow trial at 25 Hz in a PC-assisted rig: the ECP5 gate approved every commanded target, and a trusted PC/UART bridge carried the commands and issued the hold. No certification, performance level or SIL is claimed.

## Result

Run `2026-09-26-pc-timeout-stop-001` has a completed certification-folder manifest. The ECP5 approved the exact commanded targets before they were forwarded through the PC and the official LeRobot API. Once measured motion was established, a deliberately CRC-invalid UART request produced no reply. The host's receive timeout fired and it issued one separately authorized measured-position hold, keeping torque on.

Every predeclared stopping-response check passed:

| Measurement | Observed | Predeclared limit |
|---|---:|---:|
| Invalid request to timeout detection | 45.83 ms | 45–60 ms |
| Detection to hold transmission | 4.29 ms | 50 ms |
| Invalid request to hold transmission | 50.12 ms | 100 ms |
| First sampled position within the stable remainder of the trace | 79.14 ms after request | 250 ms; span ≤ 2 counts for ≥ 1 s |
| Additional sampled travel | 9 encoder counts | 16 counts |

350 samples were retained. There were two normal goal writes and one hold, with no later normal write. Torque stayed enabled on all six joints, no servo status errors were reported, and temperatures were 52–55 °C. The 50.12 ms figure includes the receive timeout; it is an observed value, not a worst-case bound.

## Provenance and retention

[Public run summary and hashes](pc-assisted-arm-timeout-2026-09-26.json) bind the completed manifest, frozen source archive, bitstream and original movie. Full source and build inputs, raw traces and video are retained under the run ID and are available to evaluators under NDA. The FPGA was configured in volatile SRAM only.