Mean triangle balance exceeded the sign-shuffle expectation for strong edges in 1–80 Hz diagonal CAR models from 36 SEEG patients

We tested signed triangle balance in public SEEG data from 36 epilepsy patients. The source was OpenNeuro ds004100 version 1.1.3. Each patient supplied two runs and four 30-second blocks. We kept contact variances and set covariance between different input contacts to zero. We then applied common average reference (CAR). In the primary 1–80 Hz complete graphs, mean balance was 1.48 percentage points below the exact edge-sign shuffle expectation. The descriptive 95 percent patient interval was -2.53 to -0.35 points. The fixed strongest-10-percent edge rule gave an excess of 29.02 points, with an interval of 21.89 to 36.14 points. Thus, edge selection changed the direction of the mean score in a model with zero input covariance between contacts. A four-contact proof shows that all triangles can be balanced under this model. These population models do not estimate neural coupling or an artifact fraction. The package preserves all 30 endpoints, covariance data, code, hashes, figures, and reproduction instructions. Two pinned raw-data checks rebuilt 12 covariance matrices exactly. A separate run matched all numerical tables and saved bootstrap arrays.

signed networks · triangle balance · common average reference · SEEG · intracranial EEG · edge-sign shuffle · covariance model · OpenNeuro