# results_moe_pstay_proxy

Reselect the same causal MoE fuse↔`linear_hp30` gate grid as `results_moe_fuse.md`, but choose the gate by **mean inner Acc (then H) at lower HMM `p_stay`** as a LODO real-switch proxy. Final scores still use challenge preset **0.994 / 0.8**. Never AASD/Carta labels for selection.

Code: `moe_pstay_proxy.py`. JSON: `results/research/moe_pstay_proxy.json`. Cache: `models/causal_moe_fuse/cache/` (CPU reuse). Log: `experiments/moe_pstay_proxy_console.log`.

**Not a repeat of** Acc@0.994-only MoE selector (`results_moe_fuse.md`) or emission-LL gates (`domain-emis-ll`).

## Method

1. Same 21 pre-specified gates (`const_fuse` / `const_hp30` / abs / ewma_abs / disagree / hard_abs / abs_and_disagree).
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 (emissions fit once per cell).
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)

| Rule | Acc@0.95 / H | Acc@0.99 / H | Acc@0.994 / H |
|---|---:|---:|---:|
| **`const_fuse` (SELECTED @ all three)** | **91.07 / 0.9466** | **91.12 / 0.9455** | **91.13 / 0.9453** |
| `ewma_abs_fuse\|t0.5\|s4` | 90.90 / 0.9453 | 90.96 / 0.9444 | 90.96 / 0.9442 |
| `abs_fuse\|t0.5\|s4` | 90.21 / 0.9423 | 90.28 / 0.9418 | 90.29 / 0.9415 |
| `hard_abs\|t0.5` | 90.26 / 0.9428 | 90.30 / 0.9420 | 90.31 / 0.9417 |
| `ewma_disagree\|t1.5\|s4` (best disagree) | 70.59 / 0.8096 | 71.41 / 0.8143 | 71.67 / 0.8153 |
| `const_hp30` | 67.22 / 0.7730 | 67.94 / 0.7784 | 68.14 / 0.7787 |

Ranking is **identical** across p_stay: pure fuse wins; mixing toward hp30 still hurts LODO inner Acc even under shorter synthetic dwell. Acc@0.95 does not promote hp30-favoring gates.

## LODO (preset 0.994/0.8) — 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 prior MoE / IRM↔SWA fuse.

## External (preset)

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

Mean gate weight `g=1.0` on both externals (pure fuse). AASD Acc **50.77** vs linear_hp30 **53.12** (Δ −2.35); far from Acc>60.

## Targets

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

## Verdict

Lower-`p_stay` inner Acc as a real-switch proxy **fails to change the MoE selector**: Acc@0.95 / Acc@0.99 / Acc@0.994 all pick **`const_fuse`**. LODO synthetic switches (even at 0.95) still reward fuse over hp30, so the gate cannot import hp30’s external Acc. Same external conclusion as `results_moe_fuse.md`. Do not rerun this Acc@lower-p_stay reselect on the same fuse↔hp30 gate grid.

## NEXT IDEA

Constrained select among **mixing-only** gates (drop `const_*`) by Acc@0.95 → forced `ewma_abs_fuse|t0.5|s4` and score AASD (tests whether the proxy helps once pure fuse is disallowed); or **online-moment on the fuse stream** (not hp30); or wait for **domain-emis-ll** emission-LL results. Avoid another Acc@{0.95,0.99,0.994} sweep of this grid.
