Guarded difference coding increased mean squared reader error by 79% at 32 bits per update on 100 later passages in nine SEEG listeners with correct packet delivery

We tested guarded difference coding for a fixed SEEG sound reader at 32 payload bits per update. Calibration used 54 early passages. The test used 100 later passages from nine people. The same later cohort was used by earlier studies. With correct packet delivery, difference coding increased equal-person mean squared reader error by 79.3%. All nine people had higher error. The paired increase was 0.00080921, with a 95% whole-person bootstrap interval of 0.00016865 to 0.00154909. The sound target error had no clear change. Guarded change ranges were wider than absolute ranges in 65 of 72 early streams. The package gives a conditional error proof, a range criterion, separate packet receivers, three figures, and all code and inputs. We downloaded the pinned source ranges again. All 154 feature files matched exactly. A fresh offline run also matched all analysis outputs. The evidence applies to this feature, reader, guard rule, and packet model. It does not establish a clinical benefit or a general limit on difference coding.

SEEG · brain-computer interfaces · difference coding · feature quantization · packet coding · signal processing