# results_decoder_sw3

Retrain published **vlaai_ws** reconstruction decoder with **denser synthetic attention switches** (seg_switch=0.75, seg_switch2=0.6, seg_post=2.5); rebuild SW2-style fuse (new decoder ± published clf-sw2 a=0.5). **Not** SW2↔hp30/swmix soft blend of old evidence.

HMM preset **0.994 / 0.8**. Code: `decoder_sw3.py`. JSON: `results/research/decoder_sw3.json`.

## LODO (new decoder × published clf-sw2)

Mean H **0.7894** Acc **71.88**

| Held-out | H | Acc % |
|---|---:|---:|
| AVGC | 0.7262 | 68.39 |
| DTU | 0.8322 | 73.72 |
| KUL | 0.7974 | 72.55 |
| UHD | 0.7791 | 72.42 |
| USTC | 0.8122 | 72.31 |
| **mean** | **0.7894** | **71.88** |

## External (preset)

| Set | H / Acc | vs Acc leader 53.53 / Carta hp30 60.54 | linear_hp30 |
|---|---|---|---|
| AASD | 0.6381 / **52.37** | ΔAcc -1.16 | 0.6422/53.12 |
| Carta2026 | 0.6493 / **58.67** | ΔAcc -1.87 | 0.6520/60.54 |

## Targets

| Target | Result |
|---|---|
| AASD Acc vs leader 53.53 | Acc **52.37** (Δ-1.16); H 0.6381 |
| Carta vs linear_hp30 60.54 | Acc **58.67**; H 0.6493 |
| LODO mean | H 0.7894 Acc 71.88 |
| AASD Acc>60? | NO |

## Decoder-alone (seed3 LODO mean from run_experiment)

Published decoder ~H 0.76. Denser-switch seed3 MEAN Acc **40.53** H **0.412** (acc1 0.463) — reconstruction collapsed under dense multi-switch contrastive. Fuse Acc is carried by published clf-sw2.

## Verdict

Denser-switch decoder fuse AASD Acc 52.37 vs leader 53.53 (Δ-1.16). Acc≪60; denser synth switches on reconstruction decoder did not lift real-switch Acc (hurt decoder-alone).

## NEXT IDEA

Do not re-grid denser seg_switch on this vlaai_ws recipe. Prefer a qualitatively new Acc frame path (e.g. band-limited TRF Path A, or a non-vlaai reconstruction target family) — not SW2↔hp30/swmix evidence blends. Band-limited env fallback unused (primary denser retrain completed).
