# results_subj_cond_calib

**Claim:** `subj-cond-calib` — subject-conditional affine/Platt on frozen Acc tip `hard_abs_swmix|t1` (+ SW2). Per LODO-train subject fit (a_s,b_s); ridge map from unlabeled logit moments → (a,b); apply on held-out / AASD / Carta subjects without test labels. CPU. Backup after spectrogram-vit-acc chance-abandon.

**DISTINCT** from `platt-acc-tip` (global only), `online-moment-sw2`, `subj-adapt-emis-tip` (HMM μ/σ), `focal-synth-adapt-readout` (stem fuse).

HMM **0.994/0.8**. Code: `subj_cond_calib.py`. JSON: `results/research/subj_cond_calib.json`.

## Selection (mean LODO Acc then H; never AASD/Carta)

**SELECTED `sw2_ridge_ab`** — mean LODO H **0.7525** Acc **72.58**

| Held-out | H | Acc % |
|---|---:|---:|
| AVGC | 0.6917 | 67.17 |
| DTU | 0.8528 | 80.59 |
| KUL | 0.7506 | 72.54 |
| UHD | 0.6865 | 70.83 |
| USTC | 0.7808 | 71.76 |
| **mean** | **0.7525** | **72.58** |

### Variant table

| Variant | LODO H / Acc | AASD H / Acc | ΔAcc vs 53.53 | Carta H / Acc |
|---------|--------------|--------------|---------------|---------------|
| `sw2_ridge_ab` **←** | 0.7525 / 72.58 | 0.6392 / 53.01 | -0.52 | 0.6513 / 59.46 |
| `tip_ridge_ab` | 0.7575 / 72.56 | 0.6371 / 53.09 | -0.44 | 0.6523 / 59.66 |
| `tip_hier_platt` | 0.7493 / 72.48 | 0.6385 / 53.01 | -0.52 | 0.6497 / 59.63 |
| `sw2_hier_platt` | 0.7461 / 72.38 | 0.6372 / 52.98 | -0.55 | 0.6534 / 59.45 |
| `tip_ridge_a` | 0.7926 / 72.23 | 0.6419 / 53.42 | -0.11 | 0.6448 / 59.76 |
| `sw2_ridge_a` | 0.7929 / 72.23 | 0.6433 / 53.47 | -0.06 | 0.6468 / 59.65 |
| `tip_raw` | 0.7916 / 72.19 | 0.6424 / 53.53 | +0.00 | 0.6460 / 59.71 |
| `sw2_raw` | 0.7914 / 72.09 | 0.6439 / 53.46 | -0.07 | 0.6470 / 59.61 |
| `sw2_global_platt` | 0.7247 / 71.49 | 0.6455 / 53.21 | -0.32 | 0.6632 / 58.96 |
| `tip_global_platt` | 0.7188 / 71.18 | 0.6440 / 53.26 | -0.27 | 0.6633 / 59.25 |
| `sw2_mean_subj_platt` | 0.6827 / 70.39 | 0.6472 / 53.23 | -0.30 | 0.6660 / 58.60 |
| `tip_mean_subj_platt` | 0.6719 / 69.93 | 0.6459 / 53.03 | -0.50 | 0.6691 / 58.78 |

## Targets

| Target | Result |
|---|---|
| AASD Acc vs tip 53.53 / hp30 53.12 | Acc **53.01** (Δ -0.52 / -0.11); H 0.6392 |
| Carta vs linear_hp30 60.54 | Acc **59.46** (Δ -1.08); H 0.6513 |
| Acc>60? | **NO** (AASD Acc 53.01) |
| Tip moved? | **NO** |

## Verdict

Subject-conditional ridge/mean/hier Platt on Acc tip does not beat hard_abs_swmix|t1 tip Acc 53.53 under honest LODO Acc/H selection.

## Do not repeat

Do not re-grid subject-conditional ridge/Platt on the same Acc tip SW2 / `hard_abs_swmix|t1` streams. Do not re-grid SW2↔swmix soft/hard. SpecViT Mel-STFT Acc front-end already chance-abandoned.

## NEXT IDEA

Unused Acc-first: **speech-envelope TRF residual into published SW2 vlaai** (identity-at-0 residual Path B that already learns) **or** **cross-dataset prototype memory** (LODO-train subject prototype bank → nearest-prototype emission shift) — not SSL/DANN/Path A classical/switch-schedule/Mixup/tip PP/SpecViT re-grid.
