# results_conf_gate_xstream

**Claim:** `conf-gate-xstream` — apply fuse→train-hp30 σ (and soft residual toward that adapted stream) **only** when causal EWMA `|fuse|` is low; else keep raw fuse.

**Honesty:** Past-only EWMA `|fuse|` (τ_gate=4 s). Adapted expert = fuse scaled to unlabeled train **hp30 σ** (τ=60 s), not raw hp30 mix. Soft residual: `fuse + (1−g)·α·(fuse_xσ−fuse)` with `g=σ(s·(ewma−thr))`. Select by mean LODO Acc then H among gated variants only. **No labels**, no entropy-min. Distinct from always-on `cross-stream-scale`, MoE fuse↔hp30 soft mix, and same-stream `online-moment-fuse`.

**Compute:** CPU on `models/causal_moe_fuse/cache/fuse_w05__*.pkl` + hp30 loaders. No stem re-infer / no GPU.

## Method

1. Load fuse held-*/all5 caches and linear_hp30 streams.
2. Online σ match τ=60 s; confidence gate τ=4 s; sharp=4; scale clip [0.5, 2].
3. Pre-specified gated variants (thr ∈ {0.5,1.0,1.5}, α ∈ {0.5,1.0}, hard/soft × xσ/resid).
4. Also score `fuse_w05` (sanity) and `always_xσ` (always-on reference; **not** selectable).
5. **Select** by mean LODO Acc then H among gated only — never AASD/Carta.

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

## LODO (preset 0.994/0.8; selection)

| Variant | mean H / Acc | AVGC H / Acc | selectable |
|---------|--------------|--------------|------------|
| fuse_w05 | 0.8725 / 82.58 | 0.8100 / 75.14 | no |
| always_xσ | 0.8139 / 82.13 | 0.7703 / 73.40 | no |
| hard_xσ_t0.5 | 0.8690 / 82.40 | 0.8081 / 74.77 | yes |
| hard_xσ_t1.0 | 0.8552 / 82.08 | 0.7902 / 73.87 | yes |
| hard_xσ_t1.5 | 0.8449 / 81.95 | 0.7827 / 73.56 | yes |
| soft_xσ_t0.5_s4 | 0.8691 / 82.50 | 0.8074 / 74.71 | yes |
| soft_xσ_t1.0_s4 | 0.8603 / 82.34 | 0.7965 / 74.21 | yes |
| hard_resid_a0.5_t0.5 | 0.8708 / 82.53 | 0.8090 / 74.96 | yes |
| hard_resid_a0.5_t1.0 | 0.8685 / 82.53 | 0.8061 / 74.65 | yes |
| hard_resid_a1.0_t0.5 | 0.8690 / 82.40 | 0.8081 / 74.77 | yes |
| **soft_resid_a0.5_t0.5_s4** | **0.8713 / 82.56** | **0.8102 / 74.95** | **SELECTED** |
| soft_resid_a1.0_t1.0_s4 | 0.8603 / 82.34 | 0.7965 / 74.21 | yes |

Tight thr + soft residual wins Acc among gated; closer to raw fuse than always-on xσ. Adapted fraction on LODO AVGC ≈ **4.5%**.

## 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 | 0.6155 / 50.77 | −0.17 | −2.35 | 0.5349 / 47.29 |
| always_xσ | 0.6147 / 50.29 | −0.65 | −2.83 | 0.5274 / 47.10 |
| **soft_resid_a0.5_t0.5_s4** | 0.6146 / **50.72** | **−0.22** | **−2.40** | 0.5331 / 47.18 |
| hard_xσ_t1.0 (best illegal Acc peek among gated) | 0.6177 / 50.68 | −0.26 | −2.44 | 0.5304 / 47.21 |

Selected adapted fraction on external ≈ **7.3%** — mostly identity fuse.

## Verdict

- **Selected** `soft_resid_a0.5_t0.5_s4`: AASD Acc **50.72** (Δ **−0.22** vs fuse 50.94; Δ **−2.40** vs `linear_hp30` **53.12**; Δ **+0.43** vs always-on xss 50.29; Δ **−0.25** vs omf 50.97). Gap to Acc 60 remains **~−9.3**.
- LODO mean **0.8713 / 82.56**; AVGC **0.8102 / 74.95** vs omf **0.8168 / 75.70** (below omf; ≈ raw fuse).
- Gating rescues always-on xσ’s AVGC crash, but does **not** import hp30’s AASD Acc edge — low-|fuse| frames are too sparse / wrong place for the Acc lift.

## Targets

| Target | Result |
|--------|--------|
| AVGC ≥0.85/78 | miss (0.8102/74.95); below omf 0.8168/75.70 |
| AASD Acc>60 | miss (50.72 vs hp30 53.12) |
| Carta vs linear_hp30 | away (47.18 vs 60.54) |

## Do not repeat

Do not rerun EWMA-|fuse|-gated fuse→hp30σ / soft residual (thr≈0.5–1.5, α≈0.5–1, hard/soft) at this HMM on the same MoE fuse cache.

## NEXT IDEA

Abandon framewise fuse↔hp30 / fuse↔fuse_xσ gates. Prefer (1) **ms-feat** domain adapt before frozen stem (in flight), (2) **trial-level** unlabeled assignment with enough hp30 weight on externals, or (3) change the **hp30 decision layer** (real-switch HMM / emissions) rather than another |fuse|-EWMA gate.
