# results_cross_stream_scale

**Claim:** `cross-stream-scale` — causal scale of IRM↔SWA fuse `w=0.5` online σ toward **train-calibrated linear_hp30 σ** (optional bias; optional 2-D fuse/hp30 moment align + LODO soft mix).

**Honesty:** Past-only EWMA μ/m₂ on fuse (and on hp30 for align2d). Target moments = unlabeled LODO-train / CORPUS **hp30** μ,σ — not fuse self-moments. Soft mix `g` chosen by mean LODO Acc then H only. **No labels**, no entropy-min. Distinct from same-stream `online-moment-fuse`, stream-stat MoE, and emission-LL gates.

**Compute:** CPU on `models/causal_moe_fuse/cache/fuse_w05__*.pkl` + `M.load_lodo_hp30` / `load_ext_hp30`. No stem re-infer.

## Method

1. Load fuse held-*/all5 caches and linear_hp30 streams (v4 recipe).
2. EWMA τ=60 s (`α≈0.00416` @ hop 0.25 s); scale clip [0.5, 2].
3. Pre-specified variants:
   - `fuse_xσ_t60` — per-trial scale fuse so running σ → train hp30 σ (keep fuse running mean)
   - `fuse_xμσ_t60` — full affine fuse → train hp30 μ,σ
   - `fuse_xσ_subj_t60` — scale-only with per-subject EMA carry
   - `align2d_mix_g0.50` / `g0.75` — affine-align **both** streams online to train hp30 (μ,σ), then `g·fuse_al+(1−g)·hp30_al`
4. **Select** by mean LODO Acc then H among adapt variants — never AASD/Carta.

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

Corpus moments (all5): fuse μ/σ **0.878 / 3.765**; hp30 μ/σ **0.033 / 1.435**.

## 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, all5 emis) | 0.6155 / 50.77 | −0.17 | −2.35 | 0.5349 / 47.29 |
| **fuse_xσ_t60** (selected) | 0.6147 / **50.29** | **−0.65** | **−2.83** | 0.5274 / 47.10 |
| fuse_xμσ_t60 | 0.6002 / 50.56 | −0.38 | −2.56 | 0.5552 / 48.02 |
| fuse_xσ_subj_t60 | 0.6113 / 50.38 | −0.56 | −2.74 | 0.5269 / 46.86 |
| align2d_mix_g0.50 | 0.6306 / 52.16 | +1.22 | −0.96 | 0.6153 / 55.40 |
| align2d_mix_g0.75 | 0.6148 / 51.49 | +0.55 | −1.63 | 0.5816 / 51.41 |

## 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 |
| **fuse_xσ_t60** | **0.8139 / 82.13** | **0.7703 / 73.40** |
| fuse_xμσ_t60 | 0.8606 / 78.23 | 0.8098 / 74.37 |
| fuse_xσ_subj_t60 | 0.7525 / 81.90 | 0.7698 / 73.31 |
| align2d_mix_g0.50 | 0.8450 / 75.79 | 0.7890 / 71.26 |
| align2d_mix_g0.75 | 0.8608 / 77.79 | 0.8159 / 73.89 |

Scale-to-hp30-σ wins Acc among adapt variants (82.13) but **regresses H** and AVGC vs raw fuse / online-moment-fuse. Soft-mix variants lose ~5–7 Acc on LODO (illegal external Acc peek ≈52 still short of 53.12/60).

## Verdict

- **Selected** `fuse_xσ_t60`: AASD Acc **50.29** (Δ **−0.65** vs fuse 50.94; Δ **−2.83** vs `linear_hp30` 53.12; Δ **−0.68** vs online-moment-fuse 50.97). Gap to Acc 60 remains **~−9.7**.
- LODO mean **0.8139 / 82.13**; AVGC **0.7703 / 73.40** — Acc near fuse, H/AVGC worse (DTU Acc↑/H↓ imbalance under σ shrink).
- Cross-stream σ geometry does **not** transfer the hp30 AASD Acc edge under honest LODO selection; always-on fuse→hp30σ hurts more than same-stream fuse→fuseσ.

## Do not repeat

Do not rerun EWMA τ≈60 fuse→train-hp30 σ/μ affine (trial or subject) or 2-D align+fixed-g soft mix at this HMM.

## NEXT IDEA

**Confidence-gated cross-stream residual:** apply fuse→hp30 σ (or residual `fuse + α(hp30_al − fuse_al)`) **only** when causal EWMA `|fuse|` is below a LODO-chosen threshold; else keep raw fuse. Tests whether relative scale helps only on uncertain frames — distinct from always-on cross-stream-scale, MoE `|fuse|` hard gates, and trial-mmd / emis-LL.
