# results_online_moment

**Claim:** `online-moment-hp30` — per-subject **causal-online** unlabeled moment match on `linear_hp30` (EMA affine to train μ/σ **or** online whiten). **No entropy-min. No labels.**

**Honesty:** Past-only EWMA of unlabeled decision-stream moments within each subject (trials sorted by key; state carried across trials unless `*_trial`). Affine uses train-calibrated CORPUS hp30 μ/σ. Distinct from entropy-min TTA (`results_tta.md`) and global CORAL (`results_ext_coral.md`).

**Compute:** CPU-only numpy/scipy on cached 4-seed logits. ~300 s. No `gpu.lock`.

**Selection:** LODO mean **Acc then H** only (never AASD/Carta). Fixed HMM **0.994 / 0.8**.

Artifacts: `online_moment_hp30.py`, `results/research/online_moment_hp30.json`, `models/online_moment_hp30/selection.json`, `experiments/online_moment_hp30.log`.

## LODO (selection)

| Variant | mean H / Acc |
|---------|--------------|
| **linear_hp30** (baseline) | **0.7797 / 68.15** |
| affine_tau30_scale (best adapted) | 0.7678 / 67.03 |
| affine_tau60 | 0.7209 / 63.87 |
| affine_tau30_trial | 0.6369 / 61.59 |
| affine_tau30 | 0.6300 / 61.34 |
| whiten_tau60 | 0.7147 / 63.62 |
| whiten_tau30 | 0.6349 / 61.13 |
| affine_tau15 | 0.5536 / 58.14 |
| whiten_tau15 | 0.5564 / 57.98 |
| affine_tau5 | 0.5355 / 55.85 |

**Selected:** `linear_hp30` (all online variants regress LODO Acc/H).

## External (report all; headline = selected)

| Variant | AASD H / Acc | ΔAcc vs 53.12 | Carta H / Acc |
|---------|--------------|---------------|---------------|
| **linear_hp30** (selected) | **0.6422 / 53.12** | **0.00** | **0.6520 / 60.54** |
| affine_tau60 | 0.6435 / 53.14 | +0.02 | 0.6560 / 60.69 |
| whiten_tau60 | 0.6427 / 53.11 | −0.01 | 0.6558 / 60.74 |
| affine_tau30 | 0.6424 / 53.00 | −0.12 | 0.6507 / 60.91 |
| affine_tau30_scale | 0.6403 / 52.86 | −0.26 | 0.6481 / 60.44 |
| affine_tau5 | 0.6352 / 52.28 | −0.84 | 0.6268 / 61.17 |

## Verdict

- Honest LODO selection keeps the unadapted `linear_hp30` stream → **AASD Acc 53.12 (Δ 0.00 vs baseline)**. Gap to 60 remains **~−6.9**.
- Online EMA affine/whiten **hurts LODO** (mean Acc down to 55–67). External AASD Acc stays flat even on adapted variants (best illegal peek +0.02 at `affine_tau60`).
- Bottleneck is not per-subject 1st/2nd-order drift of the 1-D hp30 decision stream (same conclusion as global CORAL).

## Do not repeat

Do not rerun per-subject / per-trial causal EMA affine-to-train or online-whiten on frozen `linear_hp30` (1-D mean-of-4 hp30) at this HMM. Global CORAL on the same stream already flat (`results_ext_coral.md`).

## NEXT IDEA

**Unlabeled emission-LL / domain-aware stream choice** (already claimed as `domain-emis-ll`): pick or blend fuse vs `linear_hp30` by which **train-fit Gaussian emissions** assign higher likelihood to the observed unlabeled evidence — orthogonal to moment matching of the stream itself. Alternate: moment/CORAL on **higher-dim stem ms features** before the head (not 1-D logits).
