# results_tf_directnet

**Backbone: TFDirect** (causal Conv1d patch + pre-norm MHSA+FFN ×4 → dense B,W,d). Path **abandoned**.

| Check | Value |
|---|---|
| swapΔ | **0.0255** (anti_err=0; zero_eeg_delta=0.0759) |
| AVGC chance-probe @50 | **0.6958** ≥ 0.685 → abort |
| LODO / AASD / Carta | not run |

Prep `eaP_n5_t60`, 10 s, sw2, random init, batch 64, d=64 heads=4 layers=4 patch_k=15. Code: `tf_directnet_lodo.py`, `decoders.TFDirect`. JSON: `results/direct_clf__eaP_n5_t60_w10_tf_direct_aug-sw2.json`.

Closes ALREADY_TRIED §C1 open slot (was “cancelled before any run”). Same chance fate as WaveNet / Mamba / EEG-TCNet / DeepConv. Do not rerun.

## NEXT IDEA

Not fuse↔hp30; not global 24-d CORAL. Prefer a front-end that already learns on AVGC (published vlaai_ws / Shallow) with a label-free AASD cue aimed at frame Acc (~0.54 → 0.60), not another random-init Transformer/TCN body.

