# results_clf_noise_aug_ft

Fine-tune published **SW2** `vlaai_ws` 10 s classifiers from held-*/all5 checkpoints with train-only **EEG noise / channel-drop / SpecAugment-style** curriculum (noise=0.25, p_chan=0.35, p_cluster=0.5, p_tilt=0.5, p_time_mask=0.5, p_freq_mask=0.5). `p_switch=0.5`, `clf_p_switch2=0.5` (= published sw2; **not** denser switch). Architecture unchanged. **Not** hardswitch/shorthold/switch-dur, **not** MixStyle, **not** tip Acc PP / emission-shift.

Prep `eaP_n5_t60`, FT lr=0.00025, 4×250 steps, seeds=[0, 1, 2, 3]. HMM p_stay 0.994 scale 0.8.

Code: `clf_noise_aug_ft_lodo.py`. JSON: `/root/disk_3tb/eeg-thao/results/research/clf_noise_aug_ft.json`. Ckpts: `models/clf_noise_aug_ft/s*/`.

## Headline Acc

Selector stream: **sw2_fuse_a0.5**

| Set | H / Acc | vs tip 53.53 / Carta hp30 60.54 |
|---|---|---|
| AASD | 0.6439 / **53.20** | ΔAcc -0.33 |
| Carta2026 | 0.6423 / **59.45** | ΔAcc -1.09 |
| LODO mean | 0.7900 / 71.74 | |

AASD Acc>60? **NO**

## Per-seed LODO + external

| Seed | LODO H/Acc | AASD H/Acc | Carta H/Acc |
|---|---|---|---|
| 0 | 0.7825/71.57 | 0.6317/52.02 | 0.6458/58.82 |
| 1 | 0.7738/70.82 | 0.6313/51.68 | 0.6363/58.34 |
| 2 | 0.7921/71.88 | 0.6318/52.00 | 0.6240/58.20 |
| 3 | 0.7750/70.94 | 0.6305/51.91 | 0.6331/58.36 |

## Equal-mean seed fuse

LODO mean **0.7892/72.30**. AASD **0.6354/52.21** (Δ-1.32). Carta **0.6322/58.85**.

## SW2-style fuse (pub decoder × noise-aug clf, a=0.5)

LODO mean **0.7900/71.74**. AASD **0.6439/53.20** (Δ-0.33). Carta **0.6423/59.45**.

## Targets

| Target | Result |
|---|---|
| AASD Acc vs tip 53.53 | Acc **53.20** (Δ-0.33) |
| AASD Acc>60? | NO |
| Carta vs linear_hp30 60.54 | Acc **59.45** |
| vs linear_hp30 AASD 53.12 | Δ+0.08 |

## Verdict

EEG noise/SpecAugment FT from published SW2: AASD Acc 53.20 vs tip 53.53 (Δ-0.33) via sw2_fuse_a0.5.

## NEXT IDEA

If Acc flat/~53: stronger EEG noise/channel-drop/SpecAug FT alone does not close the real-switch Acc gap (same envelope as switch-schedule FT). Prefer **speech-xattn-dec**: multi-head cross-attn decoder with speech-envelope tokens as KV and EEG latents as Q (end-to-end Acc LODO) — ≠ temp-attn-spk (self-attn on EEG+speech concat), ≠ speech-resid (FiLM/additive residual), ≠ aux-env-spk (MSE recon aux). Not emis-shift / tip gate / SSL / Path A / SpecViT.

## AVGC note (unused; do not burn GPU)

AVGC tip gap ≈−0.027 H: unused idea for coordinator — causal *per-dataset* temperature on frozen focal↔synth fuse (train-only T for AVGC fold only; ≠ Mixup/focal/DANN/SSL).
