# results_wavenet

**Backbone: causal WaveNet** (van den Oord gated dilated residuals; dilations 1..64; tanh⊗σ + residual/skip; dense `(B,W,d)`). Path **abandoned**.

## Outcome

| Check | Result |
|---|---|
| Swap-test | anti_err=0; swap_delta=0.0033; zero_eeg_delta=0.0037 (≈speech-only at init) |
| Chance probe (AVGC fold, 50 steps) | mean_loss **0.6929** ≥ 0.685 abort |
| LODO / AASD / Carta | not scored |

Reason: train BCE stuck at ~ln(2) after 50 steps — same failure mode as EEG-TCNet / DeepConvNet / EEGNet Path A.

## Recipe (not to rerun)

- Prep `eaP_n5_t60`, 10 s (W=320 @ 32 Hz), sw2, Fuglsang extra, random init
- `decoders.WaveNet`: channels=64, kern=3, dilations=(1,2,4,8,16,32,64), cycles=1, d_out=64, dropout=0.2 (~244k params)
- Code: `wavenet_lodo.py`; JSON: `results/direct_clf__eaP_n5_t60_w10_wavenet_aug-sw2.json`
- Log: `experiments/wavenet_lodo_console.log`

## vs linear_hp30 AASD Acc 53.12

No external score (aborted before fold train). Does not move AASD Acc.

## Do not repeat

Do not rerun this gated WaveNet stack (or trivial dilation/cycle retunes) as a DirectNet front-end.
