# results_multi_stem_attn_pool

Learnable **causal soft-attention** over frozen **focal + synth + IRM** stem logits.
Past-only EWMA(logit), EWMA(|logit|) at τ∈{1,4}s → MLP → softmax gates → weighted sum.
Gates trained with frame BCE on LODO train folds only (held-out never for fit).
HMM 0.994/0.8. CPU caches only.

**≠** fixed-w IRM↔stem / agree-gate / conf-resid / MoE fuse↔hp30 / Mixup/focal FT.

Code: `multi_stem_attn_pool.py`. JSON: `/root/disk_3tb/eeg-thao/results/research/multi_stem_attn_pool.json`.

## Selected: **soft_attn**

| Held-out | H | Acc % |
|---|---:|---:|
| AVGC | 0.8057 | 75.10 |
| DTU | 0.8949 | 88.82 |
| KUL | 0.8599 | 81.46 |
| UHD | 0.8505 | 80.03 |
| USTC | 0.9140 | 86.01 |
| **mean** | **0.8650** | **82.28** |

AVGC vs tip 0.8226/75.54: ΔH -0.0169 / ΔAcc -0.44

## External

| Set | H / Acc | vs tip / hp30 |
|---|---|---|
| AASD | 0.6196 / 50.99 | tip Acc 53.53; hp30 53.12 |
| Carta2026 | 0.5372 / 47.42 | hp30 60.54 |

## Targets

| Target | Result |
|---|---|
| AVGC ≥0.85/78 | miss — 0.8057/75.10 |
| Beat tip fuse AVGC | no (ΔH -0.0169) |
| AASD Acc>60 | NO (50.99) |

## NEXT IDEA

If flat near tip: soft-attn on frozen tip stems saturates like agree-gate/conf-resid. Prefer qualitatively new train-time evidence (DANN/VICReg/temporal contrastive/envelope-TRF Path B) or non-vlaai Acc front-end — not Mixup/focal/fixed-w re-grids.
