# results_dec_clf_disagree_gate

**Claim:** `dec-clf-disagree-gate` — causal disagree-gate on published SW2 **decoder ↔ clf** evidence (same streams as fuse_general).

**Rule:** same-sign → fuse_general a∈{0.5, 0.75}; disagree → clf-only / decoder-only / max-|logit| / soft `|logit|`-weighted mix.

**Distinct from** fuse-general-a-grid (fixed a), SW2↔swmix soft-w, tip Platt/domain-emis, IRM↔synth agree-gate.

**Honesty:** CPU caches only. Select by mean LODO Acc then H @ HMM **0.994/0.8**. Never AASD/Carta for selection. Code: `dec_clf_disagree_gate.py`. JSON: `results/research/dec_clf_disagree_gate.json`.

## Selection

**SELECTED `agree_a075_clf`** — mean LODO H **0.7961** Acc **72.63**

| Held-out | H | Acc % | agree frac |
|---|---:|---:|---:|
| AVGC | 0.7225 | 68.43 | 0.612 |
| DTU | 0.8404 | 74.28 | 0.627 |
| KUL | 0.8161 | 74.91 | 0.635 |
| UHD | 0.7820 | 72.91 | 0.610 |
| USTC | 0.8193 | 72.62 | 0.619 |
| **mean** | **0.7961** | **72.63** | |

### Variant grid

| Rule | mean H | mean Acc | selectable |
|---|---:|---:|---|
| `fuse_a05` | 0.7914 | 72.09 | no |
| `fuse_a075` | 0.7972 | 72.80 | no |
| `pure_dec` | 0.7488 | 66.82 | no |
| `pure_clf` | 0.7932 | 72.69 | no |
| `always_maxabs` | 0.7783 | 70.37 | no |
| `always_soft_abs` | 0.7855 | 71.22 | no |
| `agree_a05_clf` | 0.7956 | 72.54 | yes |
| `agree_a05_dec` | 0.7492 | 66.67 | yes |
| `agree_a05_maxabs` | 0.7766 | 69.83 | yes |
| `agree_a05_soft_abs` | 0.7856 | 71.02 | yes |
| `agree_a075_clf` **←** | 0.7961 | 72.63 | yes |
| `agree_a075_dec` | 0.7501 | 66.72 | yes |
| `agree_a075_maxabs` | 0.7771 | 69.85 | yes |
| `agree_a075_soft_abs` | 0.7866 | 71.13 | yes |

## External (preset)

| Set | selected H/Acc | tip Acc leader | SW2 a=0.5 | linear_hp30 |
|---|---|---|---|---|
| AASD | 0.6407 / **52.93** | 0.6424/53.53 | 0.6439/53.46 | 0.6422/53.12 |
| Carta2026 | 0.6474 / **59.37** | 0.6460/59.71 | 0.6470/59.60 | 0.6520/60.54 |

## Targets

| Target | Result |
|---|---|
| AASD Acc vs tip 53.53 / hp30 53.12 | Acc **52.93** (Δtip -0.60; Δhp -0.19); H 0.6407 |
| Carta vs hp30 60.54 | Acc **59.37** (Δ -1.17); H 0.6474 |
| LODO mean | H 0.7961 Acc 72.63 |
| AASD Acc>60? | NO |

## Verdict

Honest LODO Acc/H selector over dec↔clf same-sign→fuse / disagree policies does not create AASD Acc>60. Sign-gated mix of published SW2 streams stays in the Acc mid-50s portfolio.

## Do not repeat

Do not re-grid SW2 decoder↔clf same-sign→a∈{0.5,0.75} / disagree clf|dec|maxabs|soft_|logit| at this HMM under LODO Acc/H.

## NEXT IDEA

Do not re-grid SW2 decoder↔clf disagree-gates under LODO Acc/H — Acc stays mid-50s. Prefer a qualitatively new Acc front-end (clf denser switch FT already in flight) or Carta-side path that preserves hp30 Acc 60.54 without collapsing AASD. Keep tip Acc 53.53 and linear_hp30 Carta 60.54 as portfolio leaders.
