# results_clf_switch_dur_ft

Fine-tune published **SW2** `vlaai_ws` 10 s classifiers from held-*/all5 checkpoints with **longer-tailed switch duration** aug (`p_switch=0.5`, `clf_p_switch2=0.5` = same rate as sw2). In-window τ from LODO-train-only lognormal hold prior soft-mapped to longer long-side holds. Architecture unchanged.

**Not** clf-hardswitch-ft denser p_switch, **not** decoder-sw3 recon, **not** frozen-trunk GRU head, **not** tip Acc PP / multitask-on-vlaai.

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

## Headline Acc

Selector stream: **sw2_fuse_a0.5**

| Set | H / Acc | vs tip 53.53 / Carta hp30 60.54 |
|---|---|---|
| AASD | 0.6437 / **53.20** | ΔAcc -0.33 |
| Carta2026 | 0.6443 / **59.62** | ΔAcc -0.92 |
| LODO mean | 0.7887 / 72.03 | |

AASD Acc>60? **NO**

## Per-seed LODO + external

| Seed | LODO H/Acc | AASD H/Acc | Carta H/Acc |
|---|---|---|---|
| 0 | 0.7854/71.33 | 0.6333/52.02 | 0.6307/57.96 |
| 1 | 0.7764/71.24 | 0.6324/51.91 | 0.6298/58.73 |
| 2 | 0.7889/71.90 | 0.6303/51.65 | 0.6288/58.56 |
| 3 | 0.7820/71.13 | 0.6297/51.90 | 0.6369/58.73 |

## Equal-mean seed fuse

LODO mean **0.7958/72.73**. AASD **0.6349/52.26** (Δ-1.27). Carta **0.6416/59.23**.

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

LODO mean **0.7887/72.03**. AASD **0.6437/53.20** (Δ-0.33). Carta **0.6443/59.62**.

## 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 **59.62** |
| vs linear_hp30 AASD 53.12 | Δ+0.08 |

## Verdict

Longer-tailed switch-duration 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: in-window duration reweighting (like denser p_switch) does not close the real-switch gap. Prefer a qualitatively new Acc frame that is not tip PP, Acc soft-w, vlaai identity residuals, denser/duration switch re-grids, or frozen-trunk head-swap — e.g. non-vlaai Acc front-end warmed from published evidence, or AASD-like short-hold curriculum estimated without external labels.
