Reliable acknowledgments by the next output bounded local phase error on clear slots but increased bytes in seven SEEG replays with immediate delivery and 10% simulated slot loss

We tested phase message transport in seven public SEEG rest records from OpenNeuro ds006065 version 1.0.0. A frozen local reader used 100 ms of past voltage to forecast phase 50 ms ahead. Each replay had 11,600 outputs and four pairs. We erased exactly 10% of time slots as isolated 10 ms losses or 50 ms bursts. Each person had 100 loss schedules. A sparse rule sent a phase when local error exceeded 10 degrees. Reliable acknowledgments before the next output restored the 10 degree local agreement bound at every available output, under immediate forward delivery and a shared clock. Full updates also met this bound and needed no return traffic. Under bursts, acknowledgments reduced the all-output fraction above 10 degrees from 8.97% to 7.87%. The paired change was -1.097 percentage points, with a descriptive 95% person interval of -1.220 to -0.965. The six-byte acknowledgment rule used 16.27% more total bytes than full updates; its interval was 14.38% to 18.19%. A fresh full-source audit and an independent scalar and wire replay passed. The package contains voltage, models, loss masks, scores, actual messages, figures, proof, code, and reproduction steps. The cohort was reused. Loss was simulated. The local reader had 56.56 degrees of mean error to an offline reference before loss. The bound concerns agreement with that reader, not true neural phase. No physical link or stimulation was tested.

SEEG · phase forecasting · brain-computer interface · packet loss · acknowledgments · signal transport · reproducible research