# results_clf_shorthold_ft

Fine-tune published **SW2** `vlaai_ws` 10 s classifiers from held-*/all5 checkpoints with **shorter-hold switch curriculum** (`p_switch=0.5`, `clf_p_switch2=0.5` = same rate as sw2). In-window τ/gaps from LODO-train *short-half* hold prior (lower 50% ICDF of synth segment lengths) soft-mapped to short-side holds → rapid flips. Architecture unchanged.

**Not** clf-switch-dur-ft longer holds, **not** clf-hardswitch-ft denser p_switch, **not** tip-sdur gate, **not** decoder-sw3 / frozen-trunk GRU / tip Acc PP.

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_shorthold_ft_lodo.py`. JSON: `/root/disk_3tb/eeg-thao/results/research/clf_shorthold_ft.json`. Ckpts: `models/clf_shorthold_ft/s*/`.

## Headline Acc

Selector stream: **sw2_fuse_a0.5**

| Set | H / Acc | vs tip 53.53 / Carta hp30 60.54 |
|---|---|---|
| AASD | 0.6466 / **53.44** | ΔAcc -0.09 |
| Carta2026 | 0.6356 / **59.52** | ΔAcc -1.02 |
| LODO mean | 0.7898 / 71.92 | |

AASD Acc>60? **NO**

## Per-seed LODO + external

| Seed | LODO H/Acc | AASD H/Acc | Carta H/Acc |
|---|---|---|---|
| 0 | 0.7811/71.15 | 0.6350/52.33 | 0.6400/58.47 |
| 1 | 0.7783/70.83 | 0.6347/52.10 | 0.6396/59.08 |
| 2 | 0.7916/71.89 | 0.6360/52.50 | 0.6232/58.22 |
| 3 | 0.7766/71.08 | 0.6361/52.43 | 0.6316/58.42 |

## Equal-mean seed fuse

LODO mean **0.7904/72.43**. AASD **0.6389/52.79** (Δ-0.74). Carta **0.6392/59.18**.

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

LODO mean **0.7898/71.92**. AASD **0.6466/53.44** (Δ-0.09). Carta **0.6356/59.52**.

## Targets

| Target | Result |
|---|---|
| AASD Acc vs tip 53.53 | Acc **53.44** (Δ-0.09) |
| AASD Acc>60? | NO |
| Carta vs linear_hp30 60.54 | Acc **59.52** |
| vs linear_hp30 AASD 53.12 | Δ+0.32 |

## Verdict

Shorter-hold switch curriculum FT from published SW2: AASD Acc 53.44 vs tip 53.53 (Δ-0.09) via sw2_fuse_a0.5.

## NEXT IDEA

If Acc flat/~53: short-hold schedule FT (like longer-hold / denser p_switch) does not close the real-switch gap — close the switch-schedule family. Prefer a qualitatively new Acc frame that is not tip PP, Acc soft-w, vlaai identity residuals, switch-schedule re-grids, or frozen-trunk head-swap — e.g. non-vlaai Acc front-end warmed from published evidence, or subject-conditional evidence calibration without AASD labels.
