Exact exceptions reduced six-bit SEEG power error in post-stimulation windows from 21 people with early fixed ranges

We tested exact source-word exceptions under early fixed voltage ranges in public SEEG. We used all 35 released records from 21 people in OpenNeuro ds006519 version 1.0.0. A six-bit code used 63 voltage levels. One method clipped outside samples. The other sent their original float32 words. Both methods used the same levels within the range. A causal reader measured 70 to 110 Hz power. The primary post-window error score fell from 17.59% to 3.70% of early mean power. The paired difference was 13.89 points, with a descriptive 95% person bootstrap interval of 10.11 to 18.08 points. Twenty people improved and one tied. Exceptions were frequent: the combined test rate was 17.39%. Actual cost was 11.59 bits per source word, with headers and calibration. Mean person byte saving was 63.78% against float32. A full-precision clipped path retained 12.52% post-window error. Three of 1,754 channel occurrences still had exception error above 100% of early mean power. Thus, exact exceptions reduced the group score but did not certify each channel. All source hashes, independent packet checks, and a complete numerical replay passed. The package preserves code, actual packets, frame powers, audit source words, figures, provenance, and repeat instructions. The source labels are analysis windows. This test does not establish a stimulation effect, neural source, decoder accuracy, or live interface performance.