# results_domain_emis_acc_tip

Unlabeled-target **domain-aware emission LL** gate between Acc tip `hard_abs_swmix|t1` (sw2_swmix_blend; AASD Acc 53.53) and 4-seed `linear_hp30` (Carta Acc 60.54), DEC_OFFSET-aligned. **No retrain.** CPU on SW2/swmix/hp30 caches.

**DISTINCT from** `results_domain_emis.md` (IRM↔SWA fuse↔hp30), Acc soft/hard w re-grids (sw2-hp30 / tri-acc / sw2-swmix), and `online-moment-sw2`.

Evidence: `g·tip + (1−g)·hp30`. Emissions `(μ,σ)` fit on **train corpus only** (signed two-state ±μ or IID). At test: frame LL; `g = sigmoid(T·EWMA(LL_tip−LL_hp30))` or hard cum/EWMA.

HMM preset p_stay **0.994** scale **0.8**. Code: `domain_emis_acc_tip.py`. JSON: `results/research/domain_emis_acc_tip.json`.

## Gate selection (mean LODO Acc then H; never AASD/Carta)

74 candidates; select by mean LODO Acc then H.

**SELECTED `const_tip|two_state|T1|τ4`** — mean LODO H **0.7916** Acc **72.19**

| Held-out | gate H | gate Acc % | tip H/Acc | hp30 H/Acc |
|---|---:|---:|---|---|
| AVGC | 0.7159 | 67.94 | 0.7159/67.94 | 0.7201/65.57 |
| DTU | 0.8390 | 74.67 | 0.8390/74.67 | 0.8039/68.16 |
| KUL | 0.8064 | 73.41 | 0.8064/73.41 | 0.8155/70.51 |
| UHD | 0.7821 | 72.55 | 0.7821/72.55 | 0.7580/68.77 |
| USTC | 0.8147 | 72.38 | 0.8147/72.38 | 0.8002/67.76 |
| **mean** | **0.7916** | **72.19** | | |

### Candidate family peaks

| Family | Best Acc / H | Key |
|---|---|---|
| `const_tip` **←** | 72.19 / 0.7916 | `const_tip|two_state|T1|τ4` |
| `const_hp30` | 68.15 / 0.7795 | `const_hp30|two_state|T1|τ4` |
| `soft_raw` | 71.15 / 0.7868 | `soft_raw|iid|T8|τ4` |
| `soft_ewma` | 71.11 / 0.7884 | `soft_ewma|two_state|T8|τ1` |
| `hard_cum` | 71.17 / 0.7862 | `hard_cum|iid|T1|τ4` |
| `hard_ewma` | 71.21 / 0.7875 | `hard_ewma|iid|T1|τ1` |

## External (preset)

| Set | gate H/Acc | tip_only | hp30_aligned | linear_hp30 pub | tip pub |
|---|---|---|---|---|---|
| AASD | 0.6424 / **53.53** | 0.6424/53.53 | 0.6404/52.83 | 0.6422/53.12 | 0.6424/53.53 |
| Carta2026 | 0.6460 / **59.71** | 0.6460/59.71 | 0.6559/60.44 | 0.6520/60.54 | 0.6460/59.71 |

Mean tip-gate weight `g` on AASD/Carta = 1.000 / 1.000.

## Targets

| Target | Result |
|---|---|
| AASD Acc vs tip 53.53 / hp30 53.12 | Acc **53.53** (Δtip +0.00; Δhp +0.41); H 0.6424 |
| AASD Acc>60? | NO |
| Carta vs linear_hp30 60.54 | Acc **59.71** (Δ -0.83); H 0.6460 |
| LODO mean | H 0.7916 Acc 72.19 |

## Verdict

Honest LODO Acc/H selector collapses to **const_tip** (= Acc tip `hard_abs_swmix|t1`). Emission-LL mixing with hp30 loses train Acc/H; same qualitative failure mode as fuse↔hp30 domain-emis-ll, now on the Acc tip family.

## NEXT IDEA

Do not rerun framewise Gaussian emission-LL tip↔hp30 (or fuse↔hp30). Bottleneck is not unlabeled stream-likelihood choice between mid-50s Acc tip and Carta-strong hp30. Prefer a **qualitatively new** Acc front-end that clears chance and lifts AASD acc1, not another frozen-stream gate. Keep tip AASD Acc 53.53 and `linear_hp30` Carta 60.54 as portfolio leaders.
