# results_tri_acc_blend

Soft/hard **3-way** blend of Acc-relevant streams: **SW2 fuse** (README_BEST a=0.5; AASD Acc 53.46) × **technique-synth-switch / swmix** (partner in SW2↔swmix leader 53.53) × **linear_hp30** (4-seed τ=30; AASD 53.12, Carta 60.54), aligned by `DEC_OFFSET=3`. **Not** pairwise SW2↔swmix / SW2↔hp30 / TRF↔SW2 / SW2-as-5th.

HMM preset **0.994 / 0.8**. Code: `tri_acc_blend.py`. JSON: `results/research/tri_acc_blend.json`.

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

**SELECTED `soft_w=(1,0,0)`** — mean LODO H **0.7914** Acc **72.09**

| Held-out | H | Acc % |
|---|---:|---:|
| AVGC | 0.7099 | 67.79 |
| DTU | 0.8393 | 74.43 |
| KUL | 0.8038 | 73.11 |
| UHD | 0.7857 | 72.54 |
| USTC | 0.8180 | 72.56 |
| **mean** | **0.7914** | **72.09** |

### Soft-w simplex (reference; corners + selected)

| Rule | mean H | mean Acc |
|---|---:|---:|
| `soft_w=(1,0,0)` **←** | 0.7914 | 72.09 |
| `soft_w=(0.75,0.25,0)` | 0.7884 | 72.02 |
| `soft_w=(0.5,0.5,0)` | 0.7806 | 71.55 |
| `soft_w=(0.5,0.25,0.25)` | 0.7955 | 71.50 |
| `soft_w=(0.75,0,0.25)` | 0.7953 | 71.42 |
| `soft_w=(0.5,0,0.5)` | 0.7948 | 70.36 |
| `soft_w=(0,0.5,0.5)` | 0.7891 | 70.31 |
| `soft_w=(0,1,0)` | 0.7645 | 69.75 |
| `soft_w=(0,0,1)` | 0.7795 | 68.15 |

## External (preset)

| Set | selected H/Acc | sw2_only | swmix_only | hp30_only | SW2↔swmix 53.53 | linear_hp30 |
|---|---|---|---|---|---|---|
| AASD | 0.6439 / **53.46** | 0.6439/53.46 | 0.6416/53.22 | 0.6417/53.04 | 0.6424/53.53 | 0.6422/53.12 |
| Carta2026 | 0.6470 / **59.60** | 0.6470/59.61 | 0.6450/59.89 | 0.6519/60.52 | 0.6460/59.71 | 0.6520/60.54 |

## Targets

| Target | Result |
|---|---|
| AASD Acc vs SW2↔swmix 53.53 / hp30 53.12 | Acc **53.46** (Δlead -0.07; Δhp +0.34); H 0.6439 |
| Carta vs linear_hp30 60.54 | Acc **59.60** (Δ -0.94); H 0.6470 |
| LODO mean | H 0.7914 Acc 72.09 |
| AASD Acc>60? | NO |

## Verdict

Honest LODO Acc/H selector on SW2×swmix×hp30 does not create AASD Acc>60. 3-way Acc-stream mixing stays in the mid-50s; Carta stays near ~60.

## NEXT IDEA

Do not re-grid 3-way SW2↔swmix↔hp30 soft/hard under LODO Acc/H. Prefer a **qualitatively new** frame-evidence Acc path (front-end that clears chance and lifts AASD acc1) — e.g. decoder-sw3 denser-switch decoder — not another Acc-stream blend. Keep SW2↔swmix Acc 53.53 and `linear_hp30` Carta 60.54 as portfolio leaders.
