# results_irm_synth_agree_gate

**Claim:** `irm-synth-agree-gate` — causal agree-gate on frozen **IRM ↔ stem-synth** logits.

**Rule:** same-sign → mean (= w=0.5 fuse); disagree → max-|logit| **or** IRM-only **or** synth-only **or** soft `|logit|`-weighted mix.

**Distinct from** fixed-w IRM↔synth, MoE fuse↔hp30, fuse-conf-resid.

**Honesty:** CPU caches only (`models/fuse_conf_resid/cache/_tmp_{irm,synth}__*`). Select by mean LODO Acc then H @ HMM **0.994/0.8**. Never AASD/Carta for selection.

**Compute:** Cache hits; no GPU. Code: `irm_synth_agree_gate.py`. JSON: `results/research/irm_synth_agree_gate.json`.

## Selection

| Rule | Pick |
|------|------|
| mean LODO Acc then H (agree-* only) | **`agree_soft_abs`** |
| fuse_w05 sanity AVGC | **0.8192 / 75.52** (exact tip) |

Ranked Acc: agree_soft_abs ≈ agree_maxabs → agree_synth → agree_irm. Sign-agree frac ≈ **88–93%** (AVGC 0.882) → mean dominates.

## LODO (preset)

| Variant | mean H / Acc | AVGC H / Acc | selectable |
|---------|--------------|--------------|------------|
| fuse_w05_mean | **0.8790 / 83.11** | **0.8192 / 75.52** | no |
| always_soft_abs | 0.8798 / 83.09 | 0.8193 / 75.59 | no |
| always_maxabs | 0.8793 / 83.08 | 0.8186 / 75.57 | no |
| **agree_soft_abs** | **0.8790 / 83.08** | **0.8187 / 75.57** | **SELECTED** |
| agree_maxabs | 0.8789 / 83.08 | 0.8190 / 75.58 | yes |
| agree_synth | 0.8790 / 83.00 | 0.8152 / 75.18 | yes |
| agree_irm | 0.8727 / 82.54 | 0.8104 / 75.24 | yes |
| pure_synth | 0.8772 / 82.78 | 0.8118 / 74.91 | no |
| pure_irm | 0.8695 / 82.35 | 0.8095 / 75.12 | no |

Selected AVGC vs tip **0.8192/75.52**: **ΔH −0.0005 / ΔAcc +0.05**.

## External (preset)

| Variant | AASD H / Acc | ΔAcc vs fuse 50.92 | ΔAcc vs hp30 53.12 | Carta H / Acc |
|---------|--------------|--------------------|--------------------|---------------|
| fuse_w05_mean | 0.6187 / 50.92 | 0.00 | −2.20 | 0.5498 / 49.09 |
| **agree_soft_abs** | 0.6209 / **50.96** | **+0.04** | **−2.16** | 0.5528 / 49.23 |
| agree_maxabs | 0.6224 / 51.05 | +0.13 | −2.07 | 0.5541 / 49.34 |

## Targets

| Target | Result |
|--------|--------|
| AVGC ≥0.85 / Acc≥78 | **miss** (0.8187 / 75.57) |
| AASD Acc>60 | miss (50.96 vs tip SW2/swmix 53.53 / hp30 53.12) |

## Verdict

Agree-gates are **≈ identity with tip fuse**: high sign-agreement leaves mean on ~90% of frames; disagree policies move LODO/AVGC by **≪0.01 H**. Selected soft-|logit| disagree is flat vs tip (tiny Acc↑, tiny H↓). External AASD Acc stays ~51. No open target moved. Always-on soft_|logit| (diagnostic) is the only mild LODO H bump (0.8798) and still far from AVGC 0.85/78.

## Do not repeat

Do not re-grid IRM↔synth same-sign→mean / disagree maxabs|irm|synth|soft_|logit| at this HMM.

## NEXT IDEA

Stop IRM↔synth **decision-layer** post-process for AVGC (fixed-w, conf-resid, agree-gate all saturated). Prefer Acc front-end (SW2↔swmix / linear_hp30 family) or a train-time change that reshapes AVGC evidence — do **not** reopen AVGC hard curricula or gated fuse residuals.
