# results_trf_pathb_sw2

**Frozen train-only speech-envelope TRF as gated additive side features into published SW2 / SPK-init vlaai_ws (Path B).**
Ridge TRF EEG→env (max_lag=400 ms); reconstruct ŷ; TRF↔env corr-proxy features → Conv1d→speech; **gate init=0.25 (≠0)**. SPK remap `bb.*`→`bb.bb.*`.

**NOT** speech-resid FiLM@0; **NOT** Path A TRF / band / dual / dss alone; **NOT** trf-sw2-fuse evidence blend.

Swap-test delta=0.0079. Prep `eaP_n5_t60`, sw2, max 4 ep×250, batch 64. HMM p_stay 0.994 scale 0.8; also hp30.

Code: `trf_pathb_sw2_lodo.py` / `decoders.TRFPathBVLAAI`. Checkpoints: `models/trf_pathb_sw2/held-*.pt`, `/root/disk_3tb/eeg-thao/models/all5_direct_clf_w10b3_ptSPK_trfpathb_aug-sw2.pt`. JSON: `/root/disk_3tb/eeg-thao/results/direct_clf__eaP_n5_t60_w10b3_ptSPK_trfpathb_aug-sw2.json`.

## LODO (preset)

| Held-out | H | Acc % | hp30 H/Acc | vs published vlaai |
|---|---:|---:|---|---|
| AVGC | 0.7006 | 65.81 | 0.6783/63.29 | pub 0.7846/72.14 |
| DTU | 0.8103 | 71.69 | 0.7802/65.76 | pub 0.7846/72.14 |
| KUL | 0.7824 | 71.61 | 0.7808/67.91 | pub 0.7846/72.14 |
| UHD | 0.7703 | 70.97 | 0.7510/67.73 | pub 0.7846/72.14 |
| USTC | 0.7950 | 70.99 | 0.7865/66.73 | pub 0.7846/72.14 |
| **mean** | **0.7717** | **70.22** | **0.7554/66.29** | pub 0.7846/72.14; speech-resid AASD 52.87 |

## External (preset)

| Set | H | Acc % | hp30 H/Acc | Acc tip 53.53 | speech-resid | linear_hp30 |
|---|---:|---:|---|---|---|---|
| AASD | 0.6391 | 52.68 | 0.6407/52.95 | Δ-0.85 | Δ-0.19 | 0.6422/53.12 |
| Carta2026 | 0.6467 | 59.49 | 0.6617/61.05 | — | — | 0.6520/60.54 |

## Targets

| Target | Result |
|---|---|
| AVGC H>=0.85 Acc>=78 | miss (0.7006/65.81) |
| UHD H>=0.85 | miss (0.7703) |
| AASD Acc>60 with H>0.63 | miss (0.6391/52.68; hp30 52.95) |
| AASD Acc vs tip 53.53 | Δ -0.85 (hp30 Δ -0.58) |
| Acc>60? | NO |

## NEXT IDEA

If TRF Path B ≈ speech-resid/se (~52–53 AASD Acc): close envelope-TRF→vlaai Path B. Prefer **proto-mem-emis** (cross-dataset prototype-memory emission shift; ≠ domain-emis / subj-adapt-emis collapse-to-const) or speech cross-attn decoder ≠ temp-attn-spk. Do not reopen Path A TRF / tip subj-calib / SpecViT.
