# results_moe_mix_forced

Constrained **mixing-only** reselect of the causal MoE fuse↔`linear_hp30` gate grid: drop `const_fuse` / `const_hp30`, pick by mean inner Acc (then H) at lower HMM `p_stay`, then score at challenge preset **0.994 / 0.8**. Never AASD/Carta labels for selection.

Code: `moe_mix_forced.py` (reuses `moe_pstay_proxy.py` / `causal_moe_fuse.py` + cache). JSON: `results/research/moe_mix_forced.json`. Cache: `models/causal_moe_fuse/cache/` (CPU). Log: `experiments/moe_mix_forced_console.log`.

**Not a repeat of** Acc@0.994 → `const_fuse` (`results_moe_fuse.md`) or Acc@lower-p_stay **including** `const_*` (`results_moe_pstay_proxy.md`).

## Method

1. Same 21-gate family as MoE fuse, but **19 mixing gates only** (exclude `const_*`).
2. For each gate: mean Acc/H over 20 LODO (held, inner) cells at `p_stay ∈ {0.95, 0.99, 0.994}`, scale 0.8.
3. **Primary select:** max Acc@0.95 then H@0.95. Also report Acc@0.99 / Acc@0.994 winners.
4. Score selected gate at **0.994/0.8** on LODO 5 + AASD + Carta.

## Gate selection (inner only, mixing)

| Rule | Acc@0.95 / H | Acc@0.99 / H | Acc@0.994 / H |
|---|---:|---:|---:|
| **`ewma_abs_fuse\|t0.5\|s4` (SELECTED @ all three)** | **90.90 / 0.9453** | **90.96 / 0.9444** | **90.96 / 0.9442** |
| `ewma_abs_fuse\|t1\|s4` | 90.26 / 0.9411 | 90.35 / 0.9405 | 90.37 / 0.9403 |
| `hard_abs\|t0.5` | 90.26 / 0.9428 | 90.30 / 0.9420 | 90.31 / 0.9417 |
| `abs_fuse\|t0.5\|s4` | 90.21 / 0.9423 | 90.28 / 0.9418 | 90.29 / 0.9415 |
| `ewma_disagree\|t1.5\|s4` (best disagree) | 70.59 / 0.8096 | 71.41 / 0.8143 | 71.67 / 0.8153 |

Ranking identical across p_stay among mixing gates: forced winner is **`ewma_abs_fuse|t0.5|s4`** (matches prior MoE runner-up when `const_fuse` was allowed).

## LODO (preset 0.994/0.8) — headline = ewma_abs_fuse|t0.5|s4

| Held-out | moe H/Acc | fuse H/Acc | linear_hp30 H/Acc |
|---|---|---|---|
| AVGC | **0.8170 / 75.49** | 0.8100 / 75.14 | 0.7200 / 65.56 |
| DTU | 0.9033 / 89.37 | 0.9007 / 89.23 | 0.8043 / 68.21 |
| KUL | 0.8845 / 82.90 | 0.8749 / 82.20 | 0.8155 / 70.51 |
| UHD | **0.8612 / 80.77** | 0.8580 / 80.10 | 0.7591 / 68.78 |
| USTC | 0.9214 / 86.52 | 0.9188 / 86.25 | 0.7997 / 67.69 |
| **mean** | **0.8775 / 83.01** | **0.8725 / 82.58** | — |

Tiny LODO lift vs fuse (ΔH +0.0050, ΔAcc +0.43); AVGC still short of 0.85/78.

## External (preset)

| Set | moe H/Acc | fuse H/Acc | linear_hp30 H/Acc |
|---|---|---|---|
| AASD | 0.6204 / **50.86** | 0.6155 / **50.77** | **0.6422 / 53.12** |
| Carta2026 | 0.5461 / 48.60 | 0.5349 / 47.29 | **0.6520 / 60.54** |

Mean gate weight `g` (fuse): AASD **0.950**, Carta **0.912** — almost pure fuse; only light hp30 bleed when |fuse| is small.

**AASD Acc move vs priors:** moe-mix **50.86** vs fuse/`const_fuse` **50.77** (Δ **+0.09**) vs linear_hp30 **53.12** (Δ **−2.26**). Far from Acc>60.

## Targets

| Target | Result |
|---|---|
| AVGC H≥0.85 Acc≥78 | miss (0.8170/75.49); tiny toward vs fuse 0.8100/75.14 |
| UHD H≥0.85 | met (0.8612); already met by fuse |
| AASD Acc>60 with H>0.63 | miss (0.6204/50.86); H still <0.63 |
| Carta vs linear_hp30 | moved away (48.60 vs 60.54) |

## Verdict

Forcing mixing-only selects the expected runner-up **`ewma_abs_fuse|t0.5|s4`**. LODO improves slightly (~0.8775/83.01), but AASD Acc barely moves (+0.09 vs fuse 50.77) and stays ~2.3 below hp30 53.12. Soft |fuse|-EWMA gating remains too fuse-heavy on externals (`g≈0.91–0.95`) to import hp30’s real-switch Acc. Do not rerun mixing-only Acc@0.95 reselect on this same fuse↔hp30 grid.

## NEXT IDEA

Leave fuse↔hp30 **frame/soft MoE** alone. Prefer ideas that either (1) change the **hp30 expert** / decision layer for real switches, (2) **trial-level** unlabeled assignment (MMD / energy) with enough hp30 weight on AASD, or (3) adapt **ms features** (not 1-D logits) before frozen stem. Avoid another Acc@{0.95,0.99,0.994} sweep of this gate family.
