# results_irm_synth_fuse

Frozen IRM ↔ stem-synth-switch logit blend at HMM time. **No retrain.** Forwards from `models/stem_ms96_irm/` and `models/stem_ms96_synth_switch/` (held-* LODO, all5 external). Both abs-pos `LogitTransformer` via `F.load_ckpt` (`stem_ms96_irm_swa_fuse`).

Blend: `stream = w · IRM + (1−w) · synth`, w ∈ {0, 0.25, 0.5, 0.75, 1}. HMM preset p_stay **0.994** scale **0.8**. Also causal tau-30 high-pass of blended evidence (`hp30`; corpus emissions).

Code: `stem_ms96_irm_synth_fuse.py`. JSON: `results/research/stem_ms96_irm_synth_fuse.json`. Log: `experiments/stem_ms96_irm_synth_fuse.log` / `_console.log`.

## Weight selection

| Rule | w | Notes |
|---|---:|---|
| **Primary (honest)** | **0.5** | Fixed a priori; no labels used |
| Global from mean inner H | 1.0 | Collapsed to pure IRM (same optimism trap as IRM↔SWA fuse). **Not used as headline** |
| Held-out pick (illegal) | 0.25 | Analysis only; uses held-out scores |
| Oracle (illegal) | per fold | AVGC/DTU/KUL→0.25; UHD→0.5; USTC→0.0 |

## LODO (preset) — primary w=0.5

| Held-out | fuse H | fuse Acc % | synth H/Acc | IRM H/Acc | fuse hp30 H/Acc |
|---|---:|---:|---|---|---|
| AVGC | **0.8192** | **75.52** | 0.8118/74.91 | 0.8095/75.12 | 0.7949/71.89 |
| DTU | 0.9104 | 89.97 | 0.9205/89.74 | 0.8999/88.88 | 0.8535/74.99 |
| KUL | 0.8792 | 82.74 | 0.8800/82.62 | 0.8702/81.84 | 0.8622/76.41 |
| UHD | **0.8614** | **80.54** | 0.8468/79.85 | 0.8530/79.96 | 0.8209/74.68 |
| USTC | 0.9247 | 86.76 | 0.9270/86.78 | 0.9148/85.96 | 0.8432/72.92 |
| **mean** | **0.8790** | **83.11** | **0.8772/82.78** | **0.8695/82.35** | |

All-w held-out means (analysis; not selectors): w0 0.8772/82.78, **w0.25 0.8811/83.25**, w0.5 0.8790/83.11, w0.75 0.8750/82.73, w1 0.8695/82.35.

## External (preset) — primary w=0.5

| Set | fuse H | fuse Acc % | fuse hp30 H/Acc | linear_hp30 (4-seed) | synth | IRM |
|---|---:|---:|---|---|---|---|
| AASD | 0.6187 | **50.92** | 0.6039/50.93 | **0.6422/53.12** | 0.6174/50.98 | 0.6207/51.01 |
| Carta2026 | 0.5497 | 49.09 | 0.5762/50.52 | **0.6520/60.54** | 0.5516/49.50 | 0.5426/48.04 |

Across all w, AASD Acc stays **50.89–51.01**. Δ vs linear_hp30 53.12 ≈ **−2.20** (primary). No blend approaches Acc>60.

## Targets

| Target | Result |
|---|---|
| Keep synth LODO H tip | **yes** — primary mean **0.8790/83.11** > synth 0.8772/82.78 and > IRM↔SWA fuse 0.8725/82.58 |
| UHD H≥0.85 | **met** (0.8614; synth had 0.8468) |
| AVGC H≥0.85 Acc≥78 | miss (0.8192/75.52); best of IRM/synth/fuse family so far, still short |
| AASD Acc vs 53.12 | miss (50.92; Δ−2.20); flat vs synth 50.98 |

## Verdict

Honest w=0.5 **recovers UHD H≥0.85** while **keeping (slightly beating) synth LODO mean H**. AVGC gets a small Acc/H lift vs synth alone but stays far from 0.85/78. AASD Acc stays ~51 — no movement vs linear_hp30. Global-w by inner H collapses to pure IRM (ignore). Illegal held-out pick w=0.25 is the LODO tip (0.8811/83.25) but not a legal selector.

## vs parents / prior fuse

| Method | LODO mean H/Acc | AVGC H/Acc | UHD H | AASD Acc |
|---|---|---|---:|---:|
| IRM | 0.8695 / 82.35 | 0.8095 / 75.12 | 0.8530 | 51.01 |
| stem-synth-switch | 0.8772 / 82.78 | 0.8118 / 74.91 | 0.8468 | 50.98 |
| IRM↔SWA fuse w=0.5 | 0.8725 / 82.58 | 0.8100 / 75.14 | **0.8580** | 50.94 |
| **IRM↔synth fuse w=0.5** | **0.8790 / 83.11** | **0.8192 / 75.52** | **0.8614** | 50.92 |

## NEXT IDEA

Stop fixed-w IRM↔{SWA,RoPE,synth} blends — LODO saturated; external Acc frozen ~51. Prefer **non-stem Acc path** (keep `linear_hp30` / SW2↔hp30 family) or a train-time change that reshapes AASD evidence. Do not re-grid IRM/synth weights or hp30-on-blend of these two.
