# results_domain_emis

Unlabeled-target **domain-aware emission LL** gate between IRM↔SWA fuse `w=0.5` and 4-seed `linear_hp30`. **No retrain.** CPU on cached pickles (`models/causal_moe_fuse/cache/fuse_w05__*.pkl` + v4 hp30 loaders).

**Distinct from** `results_moe_fuse.md` (gates on `|fuse|` / disagreement). Here the gate is causal running **log-likelihood under train-fit Gaussian emissions**.

Evidence: `g·fuse + (1−g)·hp30`. Emissions `(μ,σ)` fit on **train corpus only** (signed two-state ±μ mixture, or IID on raw logit). At test: frame LL under each stream’s own train emissions (no labels); `g = sigmoid(T·EWMA(LL_fuse−LL_hp30))` or hard cum/EWMA pick.

HMM preset p_stay **0.994** scale **0.8**.

Code: `domain_emis_ll.py`. JSON: `results/research/domain_emis_ll.json`. Log: `experiments/domain_emis_ll.log` / `_console.log`.

## Gate selection (LODO train only)

74 candidates; select by mean inner Acc then H (never AASD/Carta).

| Candidate family | Best inner Acc / H | Notes |
|---|---|---|
| **const_fuse** | **91.13 / 0.9453** | **Selected headline** |
| soft_raw/ewma T=0.5 (iid/two_state) | ≤88.21 / 0.9292 | Mild mix toward hp30; loses LODO |
| soft T≥1 | ≤84.47 | Stronger mix → collapses Acc |
| hard_cum / hard_ewma | ≤69.03 | Effectively → hp30 |
| const_hp30 | 68.14 / 0.7787 | Floor |

Honest selector → **const_fuse**: gated evidence **identical to fuse w=0.5**. Soft emission-LL gates that admit hp30 lose train Acc/H; hard LL picks behave like const_hp30.

## LODO (preset)

| Held-out | gate H | gate Acc % | fuse H/Acc | hp30 (r0 fold) H/Acc |
|---|---:|---:|---|---|
| AVGC | **0.8100** | **75.14** | 0.8100/75.14 | 0.7200/65.56 |
| DTU | 0.9007 | 89.23 | 0.9007/89.23 | 0.8043/68.21 |
| KUL | 0.8749 | 82.20 | 0.8749/82.20 | 0.8155/70.51 |
| UHD | **0.8580** | **80.10** | 0.8580/80.10 | 0.7591/68.78 |
| USTC | 0.9188 | 86.25 | 0.9188/86.25 | 0.7997/67.69 |
| **mean** | **0.8725** | **82.58** | **0.8725/82.58** | |

## External (preset)

| Set | gate (=fuse) H/Acc | fuse H/Acc | linear_hp30 (4-seed) |
|---|---|---|---|
| AASD | 0.6155 / **50.77** | 0.6155 / 50.77 | **0.6422 / 53.12** |
| Carta2026 | 0.5349 / 47.29 | 0.5349 / 47.29 | **0.6520 / 60.54** |

Mean fuse-gate weight `g` on AASD/Carta = **1.0** (const_fuse).

## Targets

| Target | Result |
|---|---|
| AVGC H≥0.85 / Acc≥78 | miss (0.8100/75.14); same as fuse |
| UHD H≥0.85 | **met** (0.8580); same as fuse |
| AASD Acc>60 with H>0.63 | miss (0.6155/50.77); below linear_hp30 53.12 |
| Carta vs linear_hp30 | moved away (0.5349/47.29) |

## Verdict

Domain emission-LL stream choice **does not beat fuse** under honest LODO selection: grid collapses to `const_fuse` (same endpoint as stream-stat MoE). Mixing hp30 via emission LL hurts LODO Acc/H; hard LL picks → hp30 and lose train hard. Portfolio unchanged: fuse for LODO / `linear_hp30` for external. Do not rerun this fuse↔hp30 framewise emission-LL gate family.
