Under fixed prediction, signed 10 ms covariance reduced mean error by 1.89 percent for bipolar SEEG between shafts in 19 selected patients at 1–40 Hz

We tested whether signed lag covariance repeats after local contact subtraction. We used 19 selected public SEEG records from OpenNeuro ds003876 version 1.0.2. At 1 to 40 Hz and a 10 ms lag, the fixed signed model reduced mean later covariance error by 1.89 percent. Only bipolar links between different named shafts were used. Error fell in 13 patients. It rose in six patients. The patient bootstrap 95 percent interval was 0.64 to 3.26 percent under independent, exchangeable patient sampling. All ten fixed checks had positive means. Fresh pinned-source rebuilds matched all 19 covariance files exactly. An offline repeat matched all results. This targeted follow-up uses the same cohort as prior work. The result supports repeatable recorded covariance structure under the specified reference control. It does not establish neural causation, whole-process time reversibility, or clinical value. The package preserves code, inputs, source versions, one raw prefix, patient results, uncertainty draws, figures, and reproduction instructions.

stereoelectroencephalography · bipolar montage · lag covariance · signed covariance · reference control · epilepsy