Trial label shuffles gave 8.67% mean false rejection for fixed opposite predictions under fair independent pairs in 100 ds004362 designs

We tested label references for complete left and right cue pairs in OpenNeuro ds004362 version 1.0.0. We derived exact count laws for fixed binary predictions. Under fair independent pair orientations, ordinary trial label shuffles gave 8.6694% mean false rejection at a nominal 5% level. This result used fixed opposite predictions and 100 observed test designs. A test that kept the pairs gave 2.9536%. These are model probabilities. The source tables do not prove the null assumptions. We also trained a fixed EEG rule on run 4 and tested runs 8 and 12. For 100 eligible people, a decision that kept one left and one right prediction per pair raised mean accuracy from 55.9216% to 60.4340%. The mean gain was 4.5124 percentage points. Its paired person bootstrap 95% interval was [2.6424, 6.4016] percentage points. The decision needed both EEG windows and used no true test label. Visual cue responses and other task effects remain possible. The package contains lossless local signals, events, source identities, code, exact distributions, predictions, models, checks, and retrieval commands. Independent enumeration and numerical checks passed. A second analysis run reproduced all nine numerical result files.

EEG · motor imagery · paired labels · permutation tests · exact inference · exchangeability · classification · OpenNeuro