Ranges from each buffer reduced eight-bit SEEG probability changes compared with fixed early ranges in eight recordings

We tested whether a range from each completed SEEG buffer preserves a fixed decoder better than a fixed early range. The source was OpenNeuro ds004473, version 1.0.2. The test used eight clinical recordings, eight bipolar pairs per person, and 674 later trials. The fixed range was the early 99.5th percentile of absolute centered voltage. Both methods used eight bits per sample. The mean absolute change from the source probability was 0.003632 for the local range and 0.009682 for the fixed range. All eight person differences favored the local range. The paired mean difference was -0.006050. The descriptive 95% person bootstrap interval was -0.007180 to -0.004764. Each complete packet used 4,048 bytes, including its scales and header. The source int16 sample payload used 8,000 bytes. The local packets changed seven of 1,348 pooled decisions. Mean task accuracy did not improve. A clipping-only control showed that more bits did not remove all fixed-range error. The package contains exact inputs, binary test packets, code, models, predictions, figures, and pinned source retrieval recipes. Independent numerical checks, a full replay, and a fresh 16-range source audit passed. This was an offline test on a reused cohort. It did not test live acquisition, new recording days, or stimulation.

stereo-EEG · signal compression · brain-computer interface · probability fidelity · quantization · neural signals