Score bounds with exact fallback saved 31.7% of packet bytes and kept 1,348 fixed decisions in eight offline SEEG tests

We tested a voltage packet method on OpenNeuro ds004473 version 1.0.2. Eight fixed decoders scored 1,348 later event windows from eight clinical recordings. The method removed each window mean and tested 4, 6, 8, 10, and 12-bit codes. A score interval controlled the choice. If no code gave a certificate, the sender used exact native samples. Actual frame bytes fell by 31.68% on average across people (descriptive 95% interval: 29.09-34.56%). All decisions matched native decisions. Task balanced accuracy remained 70.69%. The certificate concerns decoder agreement. It does not establish task truth. The proof assumes exact arithmetic and correct shared setup. Floating-point checks found no bound violation in the recorded windows or 600 artificial cases. The package preserves native windows, source versions, range hashes, fixed models, saved packets, figures, code, and repeat instructions. This offline test did not measure live intent, hardware energy, or brain writing.

stereoelectroencephalography · brain-computer interface · voltage coding · score bounds · adaptive quantization · decision preservation