# results_ext_coral

**Claim:** `ext-coral-logits` — label-free CORAL / affine on frozen **4-seed** logits using AASD/Carta **unlabeled** moments only.

**Honesty:** Adapting with unlabeled *external* moments is stronger than pure train-only whitening. Still **no label leakage** (att unused for loss, selection, or moment fit). Distinct from entropy-min TTA (`results_tta.md`) and Platt/temp (`results_calib.md`). Distinct from train-time CORAL weight in v3 `gate_robust`.

**Compute:** CPU-only numpy on cached logits (`cache/logit_v2` + `ext_clf_*_w10b3sw2P_evidence_s*`). Did not wait on `gpu.lock` (held by moe-fuse).

## Method

1. Source: CORPUS 4-seed logits (optionally causal τ=30 s high-pass).
2. Target moments: AASD / Carta / pooled unlabeled frames (mean + cov); or train-only whiten.
3. Maps: CORAL `z'=(z−μ_s) C_s^{-1/2} C_t^{1/2} + μ_out` on 4-D, then mean-of-4; or 1-D mean/std affine on the decision stream.
4. Emissions from transformed CORPUS with **cache** synthetic-switch labels (reproduces published `linear_hp30`).
5. Score LODO + AASD + Carta at fixed HMM **0.994 / 0.8**. No Acc-based selection on AASD/Carta.

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

## External (preset HMM)

| Variant | AASD H / Acc | ΔAcc vs 53.12 | Carta H / Acc |
|---------|--------------|---------------|---------------|
| **linear_hp30** (sanity) | **0.6422 / 53.12** | 0.00 | **0.6520 / 60.54** |
| linear_mean | 0.6417 / 52.91 | −0.21 | 0.6454 / 59.28 |
| **coral4_hp30_AASD** (primary) | 0.6414 / **53.11** | **−0.01** | 0.6552 / 60.54 |
| coral4_hp30_AASD_covonly | 0.6419 / 53.09 | −0.03 | 0.6548 / 60.54 |
| coral4_hp30_Carta | 0.6404 / 53.04 | −0.08 | 0.6551 / 60.56 |
| coral4_hp30_pool | 0.6410 / 53.06 | −0.06 | 0.6540 / 60.55 |
| coral4_raw_AASD | 0.6407 / 52.88 | −0.24 | 0.6444 / 59.27 |
| coral4_hp30_whiten (train-only) | 0.6413 / **53.16** | **+0.04** | 0.6533 / 60.60 |
| affine1_hp30_AASD | 0.6409 / 53.05 | −0.07 | 0.6544 / 60.53 |
| affine1_hp30_AASD_scale | 0.6417 / 53.11 | −0.01 | 0.6519 / 60.53 |
| affine1_hp30_pool | 0.6404 / 53.03 | −0.09 | 0.6548 / 60.54 |

## LODO (fixed 0.994/0.8; not nested HMM)

| Variant | mean H / Acc |
|---------|--------------|
| linear_mean | 0.7932 / 72.68 |
| linear_hp30 | 0.7797 / 68.15 |
| coral4_hp30_AASD | 0.7803 / 68.28 |
| coral4_hp30_pool | 0.7820 / 68.13 |
| coral4_hp30_whiten | 0.7730 / 67.92 |
| affine1_hp30_AASD | 0.7803 / 68.30 |

(Nested-selected published `linear_hp30` LODO was 0.7899/69.21; fixed preset is expected lower.)

## Verdict

- **AASD Acc does not move** toward 60 (best external-moment CORAL 53.11; train-only whiten 53.16). Gap to 60 remains **~−6.9**.
- Carta stays ~60.5 (already met under `linear_hp30`).
- Second-order / affine match of 4-seed (hp30) logits onto unlabeled external moments is near-identity under this HMM; bottleneck is not raw 1st/2nd-order shift of these logits.

## Do not repeat

Do not rerun global CORAL / mean-std affine of frozen 4-seed (raw or hp30) onto AASD/Carta unlabeled moments at this HMM. Per-subject or online-causal moment match, or CORAL on higher-dim stem features, would be a different claim.

## NEXT IDEA

**Per-subject causal-online moment match** on `linear_hp30` (or IRM↔SWA↔hp30 MoE stream): EMA of unlabeled logit moments within each AASD/Carta subject, affine update after emit — still label-free, stronger than global pool CORAL, orthogonal to entropy-min TTA.
