Pre-registration: A Leakage-Audited Re-evaluation of Inner-Speech Decoding on OpenNeuro ds003626, with an External-Electrode Confound Control

This is a pre-registration, not a results paper. It fixes in advance every analysis decision for a re-evaluation of inner-speech EEG decoding on the open dataset OpenNeuro ds003626 (Nieto et al., Scientific Data 9:52, 2022; 10 participants, 136 channels). Three questions are posed: whether inner-speech classification exceeds an empirically determined chance level under block-held-out and leave-one-participant-out evaluation with all fitted transforms scoped to training folds; whether the same class labels can be decoded from the dataset's external (ocular and muscular) electrodes alone, as a confound control; and how far inner-speech accuracy falls below the pronounced-speech condition recorded in the same participants, used as a within-dataset upper bound. Preprocessing, feature sets, classifiers, fold definitions, permutation procedure, multiple-comparison correction and exclusion criteria are all specified here in advance. The author commits to reporting the outcome whichever way it falls, and to labelling as exploratory any analysis added after seeing the data. At the time of deposit, no data from ds003626 had been loaded, inspected or plotted by the author. Code: to be deposited with the results under an open licence.