#17 · ML / AI methods

Controls test whether EEG decoders capture speech, sentence context, or session artifacts

The shared thread is interrogation of what an EEG decoder actually reflects. In cross-subject imagined-speech decoding from scalp EEG, leakage-controlled evaluation on three public corpora finds that no participant pair transfers above chance. Per-layer decoding of Llama-3.1-8B from EEG during silent reading shows word identity peaking early while passage-level context rises with depth, with controls excluding session artifacts. A duration-aware RSSM world model for sleep staging learns stage dwell times and enforces minimum dwell as a hard decoding-time constraint. What is new is the explicit use of controls and structural constraints to qualify what decoded content means, rather than treating accuracy as the primary output.