The standalone experiment scaffold runs a conservative baseline: validate, normalize markers, preprocess (artifacts flagged, never silently deleted), event-aligned windowing, Welch bandpower features, then a baseline classifier.
Why leave-one-session-out
Windows from the same session are correlated. Ordinary k-fold would let near-duplicate windows land in both train and test, inflating scores. Leave-one-session-out / GroupKFold keeps every session's windows together, so the reported metric reflects generalization to an unseen session — the honest question.