# results_focal_synth_adapt_readout

**Claim:** `focal-synth-adapt-readout` — past-only per-subject/trial causal **scale / temperature / affine** (and train-fold Platt/scale) on frozen **focal-IRM ↔ synth fuse w=0.5**.

**Distinct from** `fuse-conf-resid` (gated residual net), `irm-synth-agree-gate`, `online-moment` on Acc tip / IRM↔SWA fuse, tip Acc PP. No stem BCE / AVGC curriculum.

**Honesty:** Label-free causal adapt where possible; train labels only for Platt/scale/temp fit on LODO train folds. Select mean LODO Acc then H @ HMM **0.994/0.8**. Never AASD/Carta for selection.

**Compute:** CPU; caches `models/focal_synth_adapt_readout/cache/` (synth reused from fuse_conf_resid). Code: `focal_synth_adapt_readout.py`. JSON: `results/research/focal_synth_adapt_readout.json`.

## Selection

| Rule | Pick |
|------|------|
| mean LODO Acc then H (adapt only) | **`platt_then_trial_temp_t30`** |
| Raw fuse sanity AVGC | **0.8226 / 75.54** |

Ranked Acc: platt_then_trial_temp_t30 → train_platt → trial_temp_t15 → trial_temp_t30 → subj_temp_t30 → trial_temp_t60

## LODO (preset)

| Variant | mean H / Acc | AVGC H / Acc | selectable |
|---------|--------------|--------------|------------|
| fuse_w05 | 0.8821 / 83.36 | 0.8226 / 75.54 | no |
| trial_temp_t15 | 0.8839 / 83.43 | 0.8247 / 75.67 | yes |
| trial_temp_t30 | 0.8833 / 83.43 | 0.8247 / 75.66 | yes |
| trial_temp_t60 | 0.8834 / 83.42 | 0.8242 / 75.65 | yes |
| subj_temp_t30 | 0.8831 / 83.43 | 0.8239 / 75.64 | yes |
| trial_scale_t30 | 0.8840 / 82.84 | 0.8256 / 75.81 | yes |
| subj_scale_t30 | 0.8832 / 82.19 | 0.8263 / 75.84 | yes |
| trial_affine_t30 | 0.8410 / 74.80 | 0.7963 / 72.05 | yes |
| subj_affine_t30 | 0.8316 / 73.51 | 0.8004 / 72.33 | yes |
| train_temp | 0.8821 / 83.36 | 0.8226 / 75.54 | no |
| train_scale | 0.8821 / 83.36 | 0.8226 / 75.54 | yes |
| train_platt | 0.8750 / 83.94 | 0.8128 / 76.65 | yes |
| **platt_then_trial_temp_t30** | **0.8757 / 84.00** | **0.8105 / 76.65** | **SELECTED** |

Selected AVGC vs tip **0.8226/75.54**: **ΔH -0.0121 / ΔAcc +1.11**.

## External (preset)

| Variant | AASD H / Acc | ΔAcc vs tip 51.06 | ΔAcc vs hp30 53.12 | Carta H / Acc |
|---------|--------------|-------------------|--------------------|---------------|
| fuse_w05 | 0.6208 / 51.06 | +0.00 | -2.06 | 0.5494 / 49.14 |
| **platt_then_trial_temp_t30** | **0.6243 / 50.92** | -0.14 | -2.20 | 0.5591 / 49.43 |

## Targets

| Target | Result |
|--------|--------|
| AVGC ≥0.85 / Acc≥78 | **miss** (0.8105 / 76.65) |
| AASD Acc>60 | miss (50.92 vs tip SW2/swmix 53.53 / hp30 53.12) |

## Verdict

Selected **`platt_then_trial_temp_t30`**: LODO mean **0.8757/84.00**, AVGC **0.8105/76.65** (Δtip -0.0121 H / +1.11 Acc). AASD Acc **50.92**, Carta **49.43**. Does not clear AVGC 0.85/78.

## Do not repeat

Do not re-grid past-only causal temp/scale/affine or train Platt/scale on frozen focal↔synth fuse w=0.5 at this HMM.

## NEXT IDEA

Stop decision-layer readout adapts on tip fuse (temp/scale/affine/Platt saturated like online-moment / calib). Prefer Acc front-end only, or a **train-time** change that reshapes AVGC evidence without BCE γ/λ/curriculum re-grids — e.g. AVGC-subject mixup / hard-negative mining inside the tip stem recipe, not another focal γ.

