# results_speech_adaln — causal ms-feature EWMA AdaLN / FiLM on stem_ms96

**Agent:** speech-adaln  
**Script:** `stem_ms96_speech_adaln.py`  
**Method:** Content-conditioned AdaLN (NOT subject/dataset/trial ID; NOT electrode XY / graph-adj). Causal EWMA mean+std of the 24-d multi-scale ms logit features (seed-logit speech-related dynamics) at τ∈{1 s, 4 s} @ 4 Hz → zero-init γ/β FiLM after each DilatedResidual LayerNorm and after each TransformerEncoder layer. Step-0 matches SWA stem (identity AdaLN).

**Init:** `models/stem_ms96_swa/{held-*,all5}.pt`  
**Train:** STEPS=1000 BATCH=512 AdamW 8e-4 cosine; EMA 0.995 eval; early-abort if loss≥0.685 @ step 50 (passed).  
**Score:** HMM p_stay 0.994 scale 0.8 + linear_hp30 on logits. Never AASD/Carta labels.

## Scores (H fraction / Acc %)

| Set | AdaLN H/Acc | hp30 H/Acc |
|---|---:|---:|
| AVGC | 0.8087 / 75.03 | 0.7721 / 70.34 |
| DTU | 0.8944 / 88.55 | 0.8293 / 71.45 |
| KUL | 0.8693 / 81.85 | 0.8402 / 74.47 |
| UHD | 0.8566 / 79.77 | 0.8083 / 73.03 |
| USTC | 0.9162 / 85.64 | 0.8300 / 70.95 |
| **mean** | **0.8690 / 82.17** | — |
| AASD | 0.6237 / 51.09 | 0.6045 / 51.02 |
| Carta2026 | 0.5391 / 47.81 | 0.5619 / 49.36 |

## Vs open targets / refs

| Target | Best so far | This run | Met? |
|--------|-------------|---------|------|
| AVGC ≥0.85/78 | fuse 0.8100/75.14 | 0.8087/75.03 | no (−0.0013 H vs fuse) |
| AASD Acc >60 | linear_hp30 53.12 | 51.09 (hp30 51.02) | no |
| UHD H ≥0.85 | fuse 0.8580 | 0.8566 | yes (kept) |
| LODO mean ≥0.85/78 | fuse 0.8725/82.58 | 0.8690/82.17 | yes (mean bars) |
| Carta Acc >60 | linear_hp30 60.54 | 47.81 | no under this readout |

No open target moved. External Acc stays ~stem/SWA family (~51 AASD), below `linear_hp30`.

## Artifacts

- Checkpoints: `models/stem_ms96_speech_adaln/{held-AVGC,DTU,KUL,UHD,USTC,all5}.pt`
- JSON: `results/research/stem_ms96_speech_adaln.json`
- Logs: `experiments/stem_ms96_speech_adaln.log`, `experiments/stem_ms96_speech_adaln_console.log`

## NEXT IDEA (targets still open)

Stem-family AdaLN/FiLM/ALiBi/GRU edits keep LODO≈SWA and do not lift AASD Acc. Prefer an **honest fusion of `linear_hp30` (external Acc) with fuse/IRM logits (train H)** with weights chosen only on train LODO inners — or a new **seed classifier** that sees raw EEG+speech and can move AASD frame Acc above ~0.54 (stem logits alone cannot). Avoid repeating IRM/RSC/SWA/softH/TTA/MixStyle/channel-XY/graph-adj/frozen-GRU/ALiBi/feat-AdaLN.
