# results_tsception

**Backbone: TSception** (Ding et al.; multi-scale temporal inception + spatial/channel mix). Path **abandoned**.

## Protocol

- Prep `eaP_n5_t60`, 10 s, sw2, Fuglsang extra, random init
- Swap-test before train: **Δ=0.0318** (≥~0.01 PASS); anti_err=0; zero_eeg_delta=0.0883
- AVGC chance-probe @50 steps: **mean_loss=0.6914 ≥ 0.685** → abort
- No LODO / all5 / AASD / Carta scoring

## Config

F1=32, scales=0.5/0.25/0.125 s → kerns (16,8,4) @ FS=32, pool=8, d_out=64, dropout=0.5; ~132k params; batch 64.

## Verdict

Same chance failure mode as WaveNet / Mamba / DeepConvNet / EEG-TCNet / TFDirect / xcorr. Do not rerun this TSception T/S inception stack as DirectNet front-end.

Code: `tsception_lodo.py`, `decoders.TSception`. JSON: `results/direct_clf__eaP_n5_t60_w10_tsception_aug-sw2.json`.
