# results_irm_stoch_fuse

Frozen IRM ↔ stem-stochdepth logit blend at HMM time. **No retrain.** Forwards from `models/stem_ms96_irm/` and `models/stem_ms96_stochdepth/` (held-* LODO, all5 external). Both abs-pos `LogitTransformer` via `F.load_ckpt` (`stem_ms96_irm_swa_fuse`).

Blend: `stream = w · IRM + (1−w) · stoch`, 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). Never used AASD/Carta for w selection.

Distinct from IRM↔synth / IRM↔SWA / IRM↔RoPE (new second stream: stochdepth).

Code: `stem_ms96_irm_stoch_fuse.py`. JSON: `results/research/stem_ms96_irm_stoch_fuse.json`. Log: `experiments/stem_ms96_irm_stoch_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 prior IRM↔* fuses). **Not used as headline** |
| Held-out pick (illegal) | 0.5 | Analysis only; equals primary |
| Oracle (illegal) | per fold | AVGC→0.0; DTU/UHD/USTC→0.5; KUL→1.0 |

## LODO (preset) — primary w=0.5

| Held-out | fuse H | fuse Acc % | stoch H/Acc | IRM H/Acc | fuse hp30 H/Acc |
|---|---:|---:|---|---|---|
| AVGC | 0.8112 | 75.38 | 0.8149/75.38 | 0.8095/75.12 | 0.7828/70.98 |
| DTU | 0.9030 | 89.19 | 0.9013/89.35 | 0.8999/88.88 | 0.8221/70.13 |
| KUL | 0.8694 | 81.96 | 0.8690/81.99 | 0.8702/81.84 | 0.8453/74.36 |
| UHD | **0.8560** | 80.30 | 0.8528/80.43 | 0.8530/79.96 | 0.8045/72.82 |
| USTC | 0.9213 | 86.22 | 0.9206/86.32 | 0.9148/85.96 | 0.8284/70.73 |
| **mean** | **0.8722** | **82.61** | **0.8717/82.69** | **0.8695/82.35** | |

All-w held-out means (analysis; not selectors): w0 0.8717/82.69, w0.25 0.8720/82.68, **w0.5 0.8722/82.61**, w0.75 0.8705/82.50, w1 0.8695/82.35.

## External (preset) — primary w=0.5

| Set | fuse H | fuse Acc % | fuse hp30 H/Acc | linear_hp30 (4-seed) | SW2↔swmix best |
|---|---:|---:|---|---|---|
| AASD | 0.6206 | **51.05** | 0.5952/50.81 | **0.6422/53.12** | **53.53** |
| Carta2026 | 0.5437 | 48.17 | 0.5646/49.48 | **0.6520/60.54** | 59.71 |

Across all w, AASD Acc stays **51.01–51.09**. Δ vs linear_hp30 53.12 ≈ **−2.07**; Δ vs SW2-swmix 53.53 ≈ **−2.48**. No blend approaches Acc>60.

## Targets

| Target | Result |
|---|---|
| AVGC H≥0.85 Acc≥78 | miss (0.8112/75.38); **below** IRM↔synth fuse 0.8192/75.52 (Δ−0.0080 H / −0.14 Acc) |
| UHD H≥0.85 | met (0.8560); below synth-fuse 0.8614 |
| Keep / beat IRM↔synth LODO tip | miss — mean 0.8722/82.61 ≪ 0.8790/83.11 |
| AASD Acc vs 53.53/53.12 | miss (51.05; Δ−2.48/−2.07) |

## Verdict

Honest w=0.5 is ≈ IRM↔SWA fuse / pure stochdepth LODO — **does not move AVGC** toward 0.85/78 and **loses** the IRM↔synth AVGC tip (0.8112 vs 0.8192). UHD stays ≥0.85. AASD Acc frozen ~51. Global-w→pure IRM (ignore). Do not re-grid IRM↔stochdepth weights.

## 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-stochdepth | 0.8717 / 82.69 | 0.8149 / 75.38 | 0.8528 | 51.07 |
| 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 |
| **IRM↔stoch fuse w=0.5** | 0.8722 / 82.61 | 0.8112 / 75.38 | 0.8560 | 51.05 |

## NEXT IDEA

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