# results_ms_partial_coral

**Claim:** `ms-partial-coral` — label-free CORAL / diagonal-affine / scale **only on dims 0:4** (raw 4-seed logits) inside the 24-d stem ms stack; dims 4:24 identity. Frozen `stem_ms96` SWA (+ fuse for selected).

**Layout check:** `ensemble_clf_transformer_v2.build_features` builds  
`ms = concat([D, hp60, hp120, z60, slope1s, mean2s], axis=1)` → raw seeds are indices **0:4**. Verified `max|ms[:,0:4]−base|≈0`.

**Honesty:** Unlabeled AASD/Carta *raw-4* feature moments only. No att/labels for fit or selection. Distinct from global 24-d `ms-feat-coral` (selected identity, AASD 51.08) and from 1-D logit CORAL.

**Compute:** Held `gpu.lock`. Reused `models/ms_feat_coral/cache/ms24_*.pkl`. Stem forward ~288 s total.

## Method

1. Cache/reuse 24-d ms features; fit maps on **raw-4 only**; embed as block-diag 24-d (I on 4:24).
2. Variants: identity, coral_AASD (s2t), coral_t2s_AASD, coral_whiten, affine_diag_AASD, affine_diag_AASD_scale.
3. Frozen SWA `held-*` / `all5`; select by mean LODO **Acc then H** @ 0.994/0.8.
4. Score AASD/Carta @ same preset. Fuse IRM↔SWA w=0.5 for selected + identity.

Artifacts: `ms_partial_coral.py`, `results/research/ms_partial_coral.json`, `models/ms_partial_coral/`, `experiments/ms_partial_coral.log`.

## LODO (SWA; select Acc then H)

| Variant | mean H / Acc | AVGC H / Acc |
|---------|--------------|--------------|
| identity | 0.8700 / 82.61 | 0.8015 / 74.91 |
| coral_AASD (s2t) | 0.8660 / 82.06 | 0.7979 / 74.41 |
| **coral_t2s_AASD (SELECTED)** | **0.8688 / 82.85** | **0.7935 / 75.34** |
| coral_whiten | 0.8362 / 81.56 | 0.7556 / 73.17 |
| affine_diag_AASD | 0.8673 / 82.14 | 0.7982 / 74.30 |
| affine_diag_AASD_scale | 0.8721 / 82.56 | 0.8059 / 74.84 |

Unlike global 24-d CORAL (s2t/affine destroyed LODO), raw-4 maps stay near identity. Selector picks **coral_t2s** (+0.24 Acc vs identity; H −0.0012).

## External (preset HMM; SWA all5)

| Variant | AASD H / Acc | ΔAcc vs 53.12 | Carta H / Acc |
|---------|--------------|---------------|---------------|
| identity | 0.6202 / 51.08 | −2.04 | 0.5507 / 48.68 |
| coral_AASD | 0.6198 / 51.08 | −2.04 | 0.5451 / 48.57 |
| **coral_t2s_AASD (selected)** | **0.6218 / 51.05** | **−2.07** | **0.5494 / 48.74** |
| coral_whiten | 0.6252 / 50.83 | −2.29 | 0.5347 / 47.16 |
| affine_diag_AASD | 0.6178 / 51.01 | −2.11 | 0.5472 / 48.50 |
| affine_diag_AASD_scale | 0.6184 / 51.05 | −2.07 | 0.5523 / 48.76 |
| fuse selected (t2s) | 0.6197 / 50.98 | −2.14 | 0.5479 / 48.46 |
| fuse identity | 0.6179 / 50.94 | −2.18 | 0.5461 / 48.38 |

## Verdict

- **Selected:** `coral_t2s_AASD` (raw-4 target→source CORAL).
- LODO mean **0.8688 / 82.85** (AVGC 0.7935/75.34).
- AASD Acc **51.05** vs linear_hp30 **53.12** (Δ **−2.07**); Carta **48.74**. Flat vs global-CORAL identity / SWA (~51.08). Gap to Acc>60 remains ~−9.
- Partial raw-4 CORAL is **less harmful** on LODO than global 24-d, but still **does not lift** AASD Acc. Bottleneck is not a 2nd-order shift of the raw seed channels inside the ms stack.

## Do not repeat

Do not rerun CORAL / diag-affine / scale-only of **raw-4 dims inside 24-d ms** through frozen stem_ms96 SWA onto unlabeled AASD moments at this HMM. Global 24-d already covered by `ms-feat-coral`.

## NEXT IDEA

Leave stem-feature moment matching. Prefer a **speech/EEG front-end** change that can move external Acc (new encoder / challenge-allowed dual readout: fuse for LODO, `linear_hp30` for external if format permits), or **layer-wise / online** adapt inside the dilated stem (not a static train→AASD raw-4 CORAL).
