# results_irm_swa_fuse

Weighted IRM ↔ SWA logit blend at HMM time. **No retrain.** Frozen forwards from `models/stem_ms96_irm/` and `models/stem_ms96_swa/` (held-* LODO, all5 external).

Blend: `stream = w · IRM + (1−w) · SWA`, 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_swa_fuse.py`. JSON: `results/research/stem_ms96_irm_swa_fuse.json`. Log: `experiments/stem_ms96_irm_swa_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: “inner” scores are train folds of the held-* model, so higher IRM weight looks optimistic. **Not used as headline** |
| Oracle (illegal selector) | per fold | AVGC/UHD→0.5; DTU/KUL/USTC→0.25. Analysis only |

## LODO (preset) — primary w=0.5

| Held-out | fuse H | fuse Acc % | SWA H/Acc | IRM H/Acc | fuse hp30 H/Acc |
|---|---:|---:|---|---|---|
| AVGC | **0.8100** | **75.14** | 0.8015/74.91 | 0.8095/75.12 | 0.7718/70.89 |
| DTU | 0.9007 | 89.24 | 0.9003/89.52 | 0.8999/88.88 | 0.8235/70.36 |
| KUL | 0.8749 | 82.20 | 0.8769/82.35 | 0.8702/81.84 | 0.8506/74.97 |
| UHD | **0.8580** | **80.10** | 0.8553/80.09 | 0.8530/79.96 | 0.8063/73.21 |
| USTC | 0.9188 | 86.25 | 0.9161/86.19 | 0.9148/85.96 | 0.8283/70.70 |
| **mean** | **0.8725** | **82.58** | **0.8700/82.61** | **0.8695/82.35** | |

All-w held-out means (for analysis; not selectors): w0 0.8700/82.61, w0.25 **0.8734**/82.66, w0.5 0.8725/82.58, w0.75 0.8704/82.47, w1 0.8695/82.35.

## External (preset) — primary w=0.5

| Set | fuse H | fuse Acc % | fuse hp30 H/Acc | linear_hp30 (4-seed) | SWA | IRM |
|---|---:|---:|---|---|---|---|
| AASD | 0.6179 | 50.94 | 0.5966/51.00 | **0.6422/53.12** | 0.6202/51.08 | 0.6207/51.01 |
| Carta2026 | 0.5461 | 48.38 | 0.5681/49.70 | **0.6520/60.54** | 0.5507/48.68 | 0.5426/48.04 |

Across all w, AASD Acc stays **50.9–51.1** (raw) and **~51.0** (hp30). No blend approaches linear_hp30 Acc 53.12, let alone Acc>60.

## Targets

| Target | Result |
|---|---|
| AVGC H≥0.85 and Acc≥78 | miss (0.8100/75.14); **tiny lift** vs IRM (+0.0005 H, +0.02 Acc); still open |
| UHD H≥0.85 | **met** (0.8580; +0.0027 vs SWA 0.8553); not lost |
| AASD Acc>60 with H>0.63 | miss (0.6179/50.94); Acc flat / slightly worse vs SWA/IRM; far below linear_hp30 |
| Carta vs linear_hp30 | moved away (0.5461/48.38) |

## Verdict

Honest blend (w=0.5) is a small LODO win on mean H and on AVGC/UHD vs either parent, and **keeps UHD H≥0.85**. It does **not** move AVGC to 0.85/78 or AASD Acc toward 60. Applying `linear_hp30` to blended stem logits does not help AASD Acc (still ~51). The 4-seed `linear_hp30` baseline remains the external Acc leader.
