# results_spectrogram_vit_acc

**Backbone: SpecViT** (EEG→causal Mel-STFT patches → tiny ViT → antisymmetric DirectNet 10s). Path **abandoned**.

| Check | Value |
|---|---|
| swapΔ | **0.0752** (anti_err=0; zero_eeg_delta=0.1109) |
| AVGC chance-probe @50 | **0.6962** ≥ 0.685 → abort |
| LODO / AASD / Carta | not run |

Prep `eaP_n5_t60`, 10 s, sw2, random init, batch 48, d=64 heads=4 layers=2 n_comp=8 n_fft=16 n_mels=12. Code: `spectrogram_vit_lodo.py`, `decoders.SpecViT`. JSON: `results/direct_clf__eaP_n5_t60_w10_spec_vit_aug-sw2.json`.

Same chance fate as TFDirect / WaveNet / Mamba / EEG-TCNet / DeepConv / FBCNet-spk. Do not rerun Mel/STFT SpecViT with this recipe.

## Backup

Ran **subj-cond-calib** (CPU) on frozen Acc tip streams — see `results_subj_cond_calib.md`.

## NEXT IDEA

Not SSL/DANN/Path A classical/switch-schedule/Mixup/SpecViT re-grid. Prefer speech-envelope TRF residual into published SW2 vlaai (identity@0) or cross-dataset prototype-memory emission shift.
