A fixed predictive code showed no clear mean error decrease at 32 bits per update in nine SEEG readers with intact transport

We tested a fixed predictive feature code for nine offline SEEG sound readers. The source was public OpenNeuro ds004703 v1.1.0. Six early passages per person set the reader and code. The test used 100 later passages and 11,700 reader outputs. Each update had eight four-bit values. An absolute reset occurred every ten updates. Other updates coded a residual from the decoded previous feature. At 32 payload bits per update, mean coding MSE fell from 0.001020439 to 0.001017249. This was a 0.31% decrease in the ratio of means. The equal-person mean paired change was -0.000003190. Its descriptive 95% person bootstrap interval was [-0.000084950, +0.000114688]. Seven people had less error and two had more. The test showed no clear mean decrease. It does not prove equal performance. Prediction had higher observed mean coding error than the parent reader-weight allocation code. Sound-target error did not show a clear decrease. All 154 feature files matched a fresh rebuild from pinned public source ranges. A separate scalar decoder checked all 800 packets. A separate-process repeat matched 1,632 result files exactly. The package contains input arrays, source receipts, code, packets, four figures, and repeat instructions. This is exploratory reuse of a small clinical cohort with shared sounds. The test assumes intact ordered transport and does not measure hardware power, intended speech, or stimulation.