# results_moe_fuse (causal-moe-fuse)

Causal **input-conditioned** gate between two frozen streams (legal MoE; no new net):

1. **fuse** = IRM↔SWA logit blend `w=0.5` (`models/stem_ms96_irm` + `models/stem_ms96_swa`)
2. **linear_hp30** = causal τ=30 s high-pass of the mean of 4 seed classifier logits

Evidence: `g·fuse + (1−g)·hp30`, with `g_t∈[0,1]` from causal stream stats only (|fuse|, EWMA |fuse|, |fuse−hp30|, hard threshold). Gate chosen by **mean inner Acc then H** on LODO non-held-out cells (never AASD/Carta). Optional logistic gate fit on train folds only; did not beat pre-spec on inner Acc/H.

HMM preset **0.994 / 0.8**. Code: `causal_moe_fuse.py`. JSON: `results/research/causal_moe_fuse.json`. Runners (analysis): `results/research/causal_moe_fuse_runners.json`. Cache: `models/causal_moe_fuse/cache/`. Log: `experiments/causal_moe_fuse_console.log`.

**Not a repeat of** fixed-w IRM↔SWA blend or blend+hp30 (`results_irm_swa_fuse.md`).

## Gate selection (inner only)

| Rule | mean inner H | mean inner Acc | Notes |
|---|---:|---:|---|
| **`const_fuse` (SELECTED)** | **0.9453** | **91.13** | Pure fuse; headline |
| `ewma_abs_fuse\|t0.5\|s4` | 0.9442 | 90.96 | Best mixing runner-up |
| `logistic` (per-held fit) | 0.9419 | 90.49 | Loses to pre-spec |
| `const_hp30` | 0.7787 | 68.14 | Pure hp30 |
| disagree / abs_and_disagree family | ≤0.815 | ≤71.7 | Strongly prefer hp30 → bad on LODO inner |

Honest selector collapses to **pure fuse**: any blend toward hp30 hurts mean inner Acc/H.

## LODO (preset) — headline = const_fuse ≡ fuse

| Held-out | moe (=fuse) H/Acc | linear_hp30 H/Acc |
|---|---|---|
| AVGC | **0.8100 / 75.14** | 0.7200 / 65.56 |
| DTU | 0.9007 / 89.23 | 0.8043 / 68.21 |
| KUL | 0.8749 / 82.20 | 0.8155 / 70.51 |
| UHD | **0.8580 / 80.10** | 0.7591 / 68.78 |
| USTC | 0.9188 / 86.25 | 0.7997 / 67.69 |
| **mean** | **0.8725 / 82.58** | — |

Identical to IRM↔SWA fuse `w=0.5`. AVGC and UHD unchanged vs fuse.

### Runner-up mixing (not headline; held-out analysis)

`ewma_abs_fuse|t0.5|s4` (favor fuse when EWMA |fuse| large): LODO mean **0.8775 / 83.01**, AVGC **0.8170 / 75.49**. Tiny held-out lift vs fuse, but **lost** the inner Acc/H selector. AASD Acc still ~50.9.

## External (preset)

| Set | moe (=fuse) H/Acc | fuse H/Acc | linear_hp30 H/Acc |
|---|---|---|---|
| AASD | 0.6155 / **50.77** | 0.6155 / 50.77 | **0.6422 / 53.12** |
| Carta2026 | 0.5349 / 47.29 | 0.5349 / 47.29 | **0.6520 / 60.54** |

AASD Acc unchanged vs fuse (~51) and still below linear_hp30 53.12; far from Acc>60. Carta stays on fuse side (away from hp30 Acc 60.54).

Best mixing runner on AASD Acc among top-3: `hard_abs|t0.5` → Acc **51.26** (still flat).

## Targets

| Target | Result |
|---|---|
| AVGC H≥0.85 Acc≥78 | miss (0.8100/75.14); **no move** vs fuse |
| UHD H≥0.85 | met (0.8580); already met by fuse |
| AASD Acc>60 with H>0.63 | miss (0.6155/50.77); Acc flat vs fuse |
| Carta vs linear_hp30 | moved away (same as fuse) |

## Verdict

Causal MoE gated by legal inner Acc/H **selects pure fuse**. Mixing toward `linear_hp30` is never rewarded on LODO inner folds, so the gate cannot import hp30’s external Acc. Distinct from fixed-w blend+hp30, but **same external conclusion**: AASD Acc stays ~51. Do not rerun this fuse↔hp30 causal gate family with the same inner Acc/H selector.

## NEXT IDEA

Select a gate (or stream) with a **LODO proxy of real-switch difficulty** that does not use AASD/Carta labels — e.g. inner Acc at a **lower p_stay** / shorter synthetic dwell — so hp30-like evidence can win the selector; or wait for a new front-end seed (EEGNet Path B) and MoE that seed’s hp30 with fuse under the same causal gate.
