A fixed nearest neighbor rule gave 71.0% mean accuracy for random labels in run pairs with shared EEG records in 42 ds007987 adults

A fixed EEG rule with one nearest neighbor predicted assigned record labels above one half when training and test epochs shared records. We tested 42 adults in OpenNeuro ds007987 version 1.1.0. Each person had open-eye and closed-eye records at runs 1 and 2 within one visit. Each run used an independent fair label swap. Every epoch in a record kept its assigned label. We evaluated all four bit pairs exactly. Mean accuracy was 71.0% with shared records and 50.0% with separate runs. The paired difference was 21.0 percentage points, with a conditional 95% person bootstrap interval of 18.0 to 23.8 points. The early-to-late check gave 68.6% with shared records. All five planned checks retained means above one half. An exact identity links one-neighbor scores to the source of each selected neighbor. A synthetic example gave 75.0% with eye-state features alone. Thus the control does not identify a unique record fingerprint. The package preserves code, full feature inputs, a proof, raw range provenance, scores, and reproduction instructions. Four fresh raw crops matched every saved feature array. All eight offline result files were reproduced byte for byte. The result applies to this label distribution, source crop, fixed processing, and split. It does not establish a neural mechanism or clinical performance.

EEG · validation · record reuse · nearest neighbor · label control · data leakage