# results_irm_rope_fuse

Weighted IRM ↔ RoPE logit blend at HMM time. **No retrain.** Frozen forwards from `models/stem_ms96_irm/` and `models/stem_ms96_rope/` (held-* LODO, all5 external).

Blend: `stream = w · IRM + (1−w) · RoPE`, 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).

**Distinct from** IRM↔SWA fuse (`results_irm_swa_fuse.md`).

Code: `stem_ms96_irm_rope_fuse.py`. JSON: `results/research/stem_ms96_irm_rope_fuse.json`. Log: `experiments/stem_ms96_irm_rope_fuse.log` / `_console.log`.

## Weight selection

| Rule | w | Notes |
|---|---:|---|
| **Primary (honest)** | **0.5** | Fixed a priori; no labels used |
| Global from mean inner Acc then H | **0.5** | Inner Acc: w0.5 **92.122** > w0.75 92.122 (tie-break H would prefer 0.75). Run log used H-first → reported global_w=0.75; Acc-first collapses to primary. **Headline = w=0.5** |
| Global H-first (run log) | 0.75 | LODO 0.8709/82.51 — slightly worse than primary; not used as headline |
| Never AASD/Carta | — | External unused for w |

## LODO (preset) — primary w=0.5

| Held-out | fuse H | fuse Acc % | RoPE H/Acc | IRM H/Acc | fuse hp30 H/Acc |
|---|---:|---:|---|---|---|
| AVGC | **0.8139** | **75.38** | 0.8124/75.66 | 0.8095/75.12 | 0.7745/70.93 |
| DTU | 0.8996 | 89.13 | 0.8969/89.15 | 0.8999/88.88 | 0.8250/70.59 |
| KUL | 0.8675 | 81.94 | 0.8678/81.88 | 0.8702/81.84 | 0.8478/74.63 |
| UHD | **0.8566** | **80.19** | 0.8541/80.11 | 0.8530/79.96 | 0.8031/73.09 |
| USTC | 0.9243 | 86.45 | 0.9245/86.59 | 0.9148/85.96 | 0.8316/71.19 |
| **mean** | **0.8724** | **82.62** | **0.8711/82.68** | **0.8695/82.35** | |

All-w held-out means (analysis; not selectors): from log — w0=RoPE 0.8711/82.68, w0.25 ≈0.8724/82.67, **w0.5 0.8724/82.62**, w0.75 0.8709/82.51, w1=IRM 0.8695/82.35.

vs IRM↔SWA fuse w=0.5: mean **0.8725/82.58**, AVGC 0.8100/75.14 — IRM↔RoPE ties mean H/Acc and **beats AVGC H** (+0.0039).

## External (preset) — primary w=0.5

| Set | fuse H | fuse Acc % | fuse hp30 H/Acc | linear_hp30 (4-seed) | RoPE | IRM |
|---|---:|---:|---|---|---|---|
| AASD | 0.6222 | **51.10** | 0.6015/50.99 | **0.6422/53.12** | 0.6216/51.04 | 0.6207/51.01 |
| Carta2026 | 0.5422 | 48.09 | 0.5652/49.62 | **0.6520/60.54** | 0.5392/48.06 | 0.5426/48.04 |

Across all w, AASD Acc stays **~51.0–51.1**. No blend approaches linear_hp30 Acc 53.12, let alone Acc>60. hp30-on-blend does not help AASD Acc (~51.0).

## Targets

| Target | Result |
|---|---|
| AVGC H≥0.85 and Acc≥78 | miss (0.8139/75.38); **tiny H lift** vs RoPE (+0.0015); Acc −0.28 vs RoPE; still open |
| UHD H≥0.85 | **met** (0.8566); +0.0025 vs RoPE |
| AASD Acc>60 with H>0.63 | miss (0.6222/51.10); Acc flat vs parents; **Δ −2.02 vs linear_hp30 53.12** |
| Carta vs linear_hp30 | moved away (0.5422/48.09) |

## Verdict

Honest blend (w=0.5) ≈ IRM↔SWA fuse on LODO mean, with a small AVGC H edge over both RoPE and IRM↔SWA fuse, 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. Fixed-w stem logit blends of IRM with SWA or RoPE are exhausted for the open Acc target.

## NEXT IDEA

Stop fixed-w IRM↔{SWA,RoPE} blends. Prefer a **non-stem** Acc path (keep `linear_hp30` as AASD leader) or a **train-time** change that actually reshapes AASD evidence (e.g. stem synth-switch already in flight; else subject-/trial-conditional adapter fit on LODO only). Do not re-grid IRM/RoPE/SWA weights.
