# results_phase2

**Choice: vlaai_ws 15 s window** (alone). Ledger had 3/5/10/20 s fused or trained; 15 s alone absent. ATCNet / EEG-TCNet / EEG Conformer / learned-adjacency graph also unused — picked the smallest hook.

Prep `eaP_n5_t60`, 3 blocks, sw2, partial SPK w10_b3 init (window-tied norms skipped), 4×250 steps/fold. HMM p_stay 0.994 scale 0.8. Also causal hp30 of evidence.

Code: `vlaai_ws15.py`. Checkpoints: `models/vlaai_ws15/held-*.pt`, `models/all5_direct_clf_w15b3_ptSPKpartial_aug-sw2.pt`. JSON: `/root/disk_3tb/eeg-thao/results/direct_clf__eaP_n5_t60_w15b3_ptSPKpartial_aug-sw2.json`.

## LODO (preset)

| Held-out | H | Acc % | hp30 H/Acc | vs stem_ms96 |
|---|---:|---:|---|---|
| AVGC | 0.6902 | 66.64 | 0.6893/64.99 | stem 0.7945/74.06 |
| DTU | 0.8194 | 75.78 | 0.8099/70.68 | stem 0.9061/89.85 |
| KUL | 0.7773 | 72.65 | 0.8052/70.90 | stem 0.8755/82.33 |
| UHD | 0.7665 | 72.11 | 0.7641/69.95 | stem 0.8407/79.47 |
| USTC | 0.7930 | 71.93 | 0.8002/68.55 | stem 0.9219/86.36 |
| **mean** | **0.7693** | **71.82** | **0.7738/69.01** | stem 0.8677/82.42 |

## External (preset)

| Set | H | Acc % | hp30 H/Acc | linear_hp30 | stem_ms96 |
|---|---:|---:|---|---|---|
| AASD | 0.6290 | 51.48 | 0.6310/52.02 | 0.6422/53.12 | 0.6158/51.54 |
| Carta2026 | 0.6245 | 58.13 | 0.6526/60.43 | 0.6520/60.54 | 0.5695/51.03 |

## Targets

| Target | Result |
|---|---|
| AVGC H>=0.85 Acc>=78 | miss (0.6902/66.64) |
| UHD H>=0.85 | miss (0.7665) |
| AASD Acc>60 with H>0.63 | miss (0.6290/51.48; hp30 Acc 52.02) |
| Carta H>0.63 Acc>60 | raw miss Acc; **hp30** H 0.6526 Acc 60.43 ≈ linear_hp30 (already MET under linear_hp30) |
| LODO mean | below stem (0.7693/71.82 vs 0.8677/82.42) |

