# results_fuse_conf_resid

**Claim:** `fuse-conf-resid` — tiny confidence-gated residual on frozen **IRM↔synth fuse w=0.5**. Residual (affine / 1-layer MLP / compact GRU, zero-init) fit on LODO **train** folds only; apply `fuse + α·resid` **only when** past-only EWMA `|fuse|` (τ=4 s) is below thr; else raw fuse.

**Distinct from** `conf-gate-xstream` (that gated fuse→hp30σ). Residual stays on the fuse stream — no hp30 mix.

**Honesty:** Train labels only on LODO train folds / all5 corpus. Select thr/α/family by mean LODO Acc then H @ HMM **0.994/0.8**. Never AASD/Carta for selection.

**Compute:** Cache hit `models/fuse_conf_resid/cache/irm_synth_w05__*.pkl` (held + all5). Residual fit CPU; no gpu.lock for scoring. Code: `fuse_conf_resid.py`. JSON: `results/research/fuse_conf_resid.json`.

## Selection

| Rule | Pick |
|------|------|
| mean LODO Acc then H (gated only) | **`affine_a1.0_t1.0`** (α=1, thr=1.0) |
| Raw fuse sanity | AVGC **0.8192/75.52**; mean **0.8790/83.11** |

Ranked Acc tip: affine_a1.0_t1.0 → affine_a0.5_t1.5 → mlp_a0.5_t1.0 → gru_a1.0_t1.0 …

## LODO (preset)

| Variant | mean H / Acc | AVGC H / Acc | selectable |
|---------|--------------|--------------|------------|
| fuse_w05 | 0.8790 / 83.11 | **0.8192 / 75.52** | no |
| always_affine_a1 | 0.8736 / 83.77 | 0.8182 / 76.58 | no |
| **affine_a1.0_t1.0** | **0.8715 / 83.66** | **0.8080 / 76.62** | **SELECTED** |
| affine_a0.5_t1.5 | 0.8787 / 83.64 | 0.8189 / 76.36 | yes |
| mlp_a0.5_t1.0 | 0.8769 / 83.60 | 0.8167 / 76.27 | yes |
| gru_a1.0_t1.0 | 0.8786 / 83.60 | 0.8162 / 76.21 | yes |

Selected gated fraction ≈ **14.5%** frames on AVGC (thr=1.0).

## External (preset)

| Variant | AASD H / Acc | ΔAcc vs fuse 50.92 | ΔAcc vs hp30 53.12 | Carta H / Acc |
|---------|--------------|--------------------|--------------------|---------------|
| fuse_w05 | 0.6187 / 50.92 | 0.00 | −2.20 | 0.5498 / 49.09 |
| **affine_a1.0_t1.0** | 0.6249 / **51.04** | **+0.12** | **−2.08** | 0.5495 / 48.88 |

Selected ext gated frac ≈ **22.5%**.

## Targets

| Target | Result |
|--------|--------|
| AVGC ≥0.85 / Acc≥78 | **miss** (0.8080 / 76.62); vs tip 0.8192/75.52 → **ΔH −0.0112 / ΔAcc +1.10** |
| AASD Acc>60 | miss (51.04 vs tip SW2/swmix 53.53 / hp30 53.12) |

## Verdict

LODO Acc selector prefers a **stronger gated affine residual** (α=1, thr=1): mean Acc **+0.55** vs raw fuse, AVGC Acc **+1.10**, but AVGC **H regresses** (−0.011) and mean H drops below the IRM↔synth tip. External AASD Acc stays ~51. Gate-on-fuse residuals trade H for Acc; they do **not** clear 0.85/78.

## Do not repeat

Do not rerun EWMA-|fuse|-gated affine/MLP/GRU residual grids on frozen IRM↔synth w=0.5 at this HMM (same thr/α family).

## NEXT IDEA

Abandon fuse-stream conf-gated resid for AVGC. Prefer a **decision-layer** change that does not carve low-|fuse| frames (e.g. real-switch / emission retune on Acc tip SW2↔swmix), or leave IRM↔synth fuse as the LODO H tip and push Acc only on the Acc front-end. Do not reopen AVGC hard curricula / loss-λ.
