A fixed SEEG reader gave 50.3% mean accuracy across formats on unseen numbers in eight patients

We tested small and large number reading in public SEEG data. The test held out both number values and stimulus format. A fixed ridge reader gave 50.29% mean accuracy across formats in eight patients with 482 passive trials. The 95% patient resampling interval was 46.42% to 53.99%. The gain against before-onset inputs was 0.57 percentage points, with interval -3.52 to 5.01 points. Of 70 balanced number-label assignments, 32 gave at least the usual small/large accuracy. These assignments contain 35 complement pairs. The result did not give clear evidence of transfer for this reader. The package supplies a reusable test, compact inputs, models, predictions, figures, source checks, and code. All 482 analyzed EDF ranges were checked and rebuilt. A fresh-input replay matched the original results exactly. Shared task orders, a known reused cohort, sparse trials, and fixed features limit the result. It does not measure intended output, online control, or brain stimulation.

stereoelectroencephalography · number cognition · cross-format transfer · held-out values · ridge regression · brain-computer interfaces