Notes

Devlog entries from ongoing work on EEG seizure prediction and wavelet coherence methods — debugging notes, ablations, and the occasional real bug, written up from session notes.

2026-08-24 A third temporal backend: Mamba, and its real cost 2026-08-24 A compact cache, a k-sweep, and a dead encoder removed 2026-08-24 dense_edge (conv) vs dense_edge_gru: a smoke-scale comparison 2026-08-24 A CWT node encoder, and what node_only actually tests 2026-08-23 A fixed-shape channel subset: full graph, live clique only 2026-08-23 A full 6-fold real run — and a real recall cost hiding in the k-of-n win 2026-08-23 Windows/CUDA bf16: an 8–11x speedup, a real OOM fix, and k-of-n smoothing 2026-08-22 truong_stft_cnn: a single-fold CUDA trace, and a real 6-fold run 2026-08-21 A dataset-baked pod image, a safer launch script, and a cache removed entirely 2026-08-20 A torch-native CWT module, a real GPU benchmark, and the bottleneck it uncovered 2026-08-19 truong_stft_cnn as a real prediction-mode comparison baseline 2026-08-17 Checking whether the prediction results are real: two leakage claims and a label-permutation control 2026-08-17 OOM fix, a streaming classifier, and the first trustworthy prediction-mode results 2026-08-08 Subject 1 vs. subject 2 on the sparse-evidence GNN: event starvation, phase-threshold sensitivity, and mu-band loss 2026-08-07 Phase-gate direction bug: one-sided gate reported as a bug, changed, reverted 2026-08-06 Accuracy regression (0.80 → 0.71) traced to a fixed receptive field