Cross-Subject Auditory MEG Decoding for Mental Imagery: A Rank-Fusion Ensemble of Subject-Aligned Linear Models

A 4th-place solution (0.192 accuracy, 173 teams) to the IMAGINE Decoding Challenge: predicting which of ten word categories a participant was imagining from MEG recordings. The final submission is a rank-fusion ensemble of eleven cross-subject classifiers — a multi-band time-domain template matcher plus ten linear logistic-regression heads trained on per-subject Euclidean-aligned, band-pass-filtered, time-windowed features, resolved by a Hungarian balanced assignment. Training leave-one-subject-out accuracy 0.172; public leaderboard 0.192. The report also documents several negative results constraining future work, including failed localizer→imagery transfer and chance-level Riemannian methods