Under a sample time assumption, a fixed SEEG power model had negative mean transfer skill for numbers 1 to 5 in six recordings

We tested number size transfer in public SEEG data from OpenNeuro ds005415 version 1.0.0. The test used six recordings. Each training and test pair changed sensory modality and symbol type. Digits were paired with beeps. Spoken numbers were paired with dots. Both directions were tested. Each pair kept four or five shared values between 1 and 5. The tests used 473 distinct trials. A fixed linear ridge model used 80 Hz to below 150 Hz power. Mean skill was -0.113 for response power change and -0.075 for prestimulus power. Skill compares number-balanced squared error with a constant-number benchmark. The response gain was -0.038. Its descriptive 95% patient bootstrap interval was [-0.113, 0.037]. The exact two-sided patient sign-flip p was 0.46875. The conditional label shuffle p was 0.836 for positive gain. All fixed checks had negative mean response skill. Fresh downloads matched all 653 raw range hashes and every feature array. A separate regression solution confirmed the predictions. A full offline repeat matched all 53 numerical output files byte for byte. The package preserves source records, features, fits, predictions, code, and reproduction commands. Results assume that the acquisition sample column uses the EDF sample time origin. The cohort is reused from prior timing work. These errors describe the tested model. They do not establish the absence of a shared number code.

number cognition · stereotactic EEG · SEEG · cross-format transfer · ridge regression · neural decoding