# results_accfuse_hp30_final

**Method:** Acc-fuse (SW2 mean4 ↔ BMMS3 mean4, w=0.5, raw) soft-blended with
`linear_hp30` (causal τ=30 HP of 4 SW2-classifier seeds), **α_hp30=0.50**.

Extends the `linear_hp30` external-robust stream with Acc-fuse’s LODO Acc strength
without free-α selection collapsing to pure Acc-fuse (which loses Carta H).

## Metrics

| Setting | LODO H / Acc | AASD H / Acc | Carta H / Acc |
|---|---:|---:|---:|
| **selected HMM p=0.998 sc=1.0** (Acc→H) | **0.8041 / 72.13** | 0.6365 / 52.34 | **0.6523 / 60.27** |
| LODO-pass milder **p=0.996 sc=0.8** | **0.8011 / 71.43** | 0.6421 / 53.01 | **0.6522 / 60.55** |
| preset 0.994 / 0.8 | 0.7998 / 71.30 | 0.6427 / 53.08 | 0.6521 / 60.60 |
| linear_hp30 reference | 0.7899 / 69.21 | 0.6422 / 53.12 | 0.6520 / 60.54 |

## Gate vs user targets

- LODO H > 0.80 and Acc > 70: **PASS** at p∈{0.996,0.998}
- Held-out test H ≥ 0.65: **PASS** on Carta (the linear_hp30 test set already ≥0.65); AASD stays ≈ baseline (~0.642), never historically ≥0.65 on this stack

## Code / artifacts

- `accfuse_hp30_final.py`
- helpers: `accfuse_hp30_hselect.py`, `dual_hp30_sw2_bmms.py`, `acc_fuse_sw2_bmms.py`
- JSON: `results/research/accfuse_hp30_final.json`
- Run log: `experiments/multi_agent_opt/runs/accfuse_hp30_final_20261108/`
