# results_online_moment_fuse

**Claim:** `online-moment-fuse` — per-trial / per-subject **causal-online** unlabeled EWMA moment match on IRM↔SWA fuse `w=0.5` (not hp30).

**Honesty:** Past-only EWMA of mean/2nd-moment on the **fuse** decision stream; causal affine (scale clipped to [0.5, 2]). Train target = unlabeled LODO-train / CORPUS μ,σ of fuse — or online whiten. **No labels**, no entropy-min. Distinct from hp30 moment match (`results_online_moment.md`), stream-stat MoE (`results_moe_fuse.md`), and emission-LL gate (`domain-emis-ll`).

**Compute:** CPU-only on `models/causal_moe_fuse/cache/fuse_w05__*.pkl`. Did not re-infer stems; did not rerun hp30 moment match.

## Method

1. Load fuse logits: held-* for LODO, all5 corpus/external for emissions + AASD/Carta.
2. EWMA τ=60 s (`α≈0.00416` at hop 0.25 s) of local μ / m₂ using frames **before** emit.
3. Affine variants (pre-specified):
   - `trial_match_t60` — per-trial match to train μ,σ
   - `subj_match_t60` — per-subject carry across trials, match train μ,σ
   - `trial_whiten_t60` — per-trial online whiten to 0/1
   - `trial_scale_t60` — per-trial variance-only toward train σ (keep running mean)
4. Emissions from adapted CORPUS with cache synthetic-switch labels (reproduces raw fuse when identity).
5. **Select** by mean LODO Acc then H among adapt variants only — never AASD/Carta.

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

## External (preset HMM 0.994 / 0.8)

| Variant | AASD H / Acc | ΔAcc vs fuse 50.94 | ΔAcc vs hp30 53.12 | Carta H / Acc |
|---------|--------------|--------------------|--------------------|---------------|
| **fuse_w05** (sanity) | **0.6179 / 50.94** | 0.00 | −2.18 | **0.5461 / 48.38** |
| trial_match_t60 | 0.6115 / 51.20 | +0.26 | −1.92 | 0.5593 / 49.49 |
| subj_match_t60 | 0.6043 / 51.08 | +0.14 | −2.04 | 0.5573 / 49.55 |
| trial_whiten_t60 | 0.6006 / 50.82 | −0.12 | −2.30 | 0.5732 / 49.34 |
| **trial_scale_t60** (selected) | 0.6183 / **50.97** | **+0.03** | **−2.15** | 0.5498 / 48.66 |

## LODO (fixed 0.994/0.8; selection criterion)

| Variant | mean H / Acc | AVGC H / Acc |
|---------|--------------|--------------|
| fuse_w05 | 0.8725 / 82.58 | 0.8100 / 75.14 |
| trial_match_t60 | 0.8569 / 78.43 | 0.7999 / 73.96 |
| subj_match_t60 | 0.8406 / 76.18 | 0.7970 / 74.31 |
| trial_whiten_t60 | 0.8542 / 76.77 | 0.8131 / 73.83 |
| **trial_scale_t60** | **0.8769 / 83.01** | **0.8168 / 75.70** |

Match/whiten hurt LODO (~−4–6 Acc). Scale-only is the LODO winner and slightly **above** raw fuse (+0.44 H / +0.43 Acc mean; AVGC +0.0068 / +0.56).

## Verdict

- **Selected** `trial_scale_t60`: AASD Acc **50.97** (Δ **+0.03** vs fuse 50.94; Δ **−2.15** vs `linear_hp30` 53.12). Gap to Acc 60 remains **~−9**.
- LODO mean **0.8769 / 83.01**; AVGC **0.8168 / 75.70** — tiny train lift, still short of AVGC bars (0.85 / 78).
- Carta stays ~48–49 under fuse family (far below hp30 Carta 60.54).
- Online 1st/2nd-moment affine on **fuse** is near-null for external Acc under honest LODO selection — same external bottleneck as on hp30, and does **not** transfer the hp30 AASD Acc edge.

## Do not repeat

Do not rerun EWMA τ≈60 causal mean/std match or whiten on the **1-D IRM↔SWA fuse w=0.5** stream (per-trial or per-subject) at this HMM. hp30 stream already covered in `results_online_moment.md`.

## NEXT IDEA

**Cross-stream causal scale:** past-only match of fuse online σ toward **train-calibrated linear_hp30 σ** (or joint 2-D (fuse, hp30) moment align before a LODO-selected soft mix) — tests whether the external Acc gap is a relative scale geometry between the two frozen streams, without labels / entropy-min. Distinct from emission-LL gate (`domain-emis-ll`) and stream-stat MoE.
