Full chain fits gave positive prediction gains for disjoint bipolar pairs with 20 ms histories in independent contact models from two SEEG records

We tested whether removing edges between channels with a common contact removes directed prediction gains. We used 85 ten-second windows from two resting SEEG records in OpenNeuro ds005670, version 1.0.0. The model kept each measured contact autocovariance and set all cross-contact covariance to zero. Adjacent contact differences formed complete bipolar chains. With two past samples at 100 Hz, all 156,094 directed pair scores across windows were positive for channels with separate contacts. Patient medians of window median gains were 0.0000335 and 0.0000142 nats. Most gains were small. Of the primary gains, 212 exceeded 0.01 nats. Selecting alternating channels before fitting gave exactly zero model gains for the same pairs. Time reversal left all model gains unchanged. Finite phase draws retained estimation error. A closed one-lag formula explains the mechanism. A complete raw rebuild matched four saved input files. A repeat matched all nine numerical result files. The package preserves the formula, covariance tool, benchmark inputs, code, source hashes, figures, and instructions. This is a model test based on two patients. It does not establish neural connections or the frequency of this effect in acquired networks.

SEEG · bipolar channels · prediction gain · signal mixing · contact autocovariance · time reversal