Anonymous Derived Entropy Features and Replication Code for Compact Spatiotemporal Graph Attention Analysis

An anonymized, English-language archive containing 26 derived one-second epoch-level entropy feature tables (15,494 rows and 180 feature columns), supplied binary state labels, a compact temporal-attention and dynamic-graph implementation, frozen configurations, leave-one-subject-out evaluation scripts, unit tests, fold-level metrics, paired statistical comparisons, and environment documentation. The public CSV tables were verified to reproduce every supplied float32 feature and label exactly. No raw EEG, raw behavioral/PVT records, direct identifiers, demographics, manuscripts, figures, credentials, or publication-identifying information are included. The supplied state blocks are almost perfectly recoverable from within-subject row position (99.9871% mean LOSO accuracy); therefore, this archive supports secondary state-block classification and confound diagnostics, not independent validation of fatigue physiology. Primary subject-wise normalization is transductive; strict training-only inductive results are reported separately.