# results_platt_acc_tip

**Claim:** `platt-acc-tip` — train-fold-only **Platt** (a·x+b) and optional **isotonic** (logit∘iso) on Acc tip streams (CPU):
1. Acc leader `hard_abs_swmix|t1` (AASD Acc tip 53.53)
2. published SW2 `fuse_general` a=0.5

**DISTINCT** from stem `results_calib.md` Platt, `online-moment-sw2`, Acc soft-w blends, and `domain-emis-acc-tip` (emission-LL).

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

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

**SELECTED `tip_raw`** — mean LODO H **0.7916** Acc **72.19**

| Held-out | H | Acc % |
|---|---:|---:|
| AVGC | 0.7159 | 67.94 |
| DTU | 0.8390 | 74.67 |
| KUL | 0.8064 | 73.41 |
| UHD | 0.7821 | 72.55 |
| USTC | 0.8147 | 72.38 |
| **mean** | **0.7916** | **72.19** |

### Selected fit (corpus / summary)

```
{
  "mode": "none"
}
```

### Variant table (LODO Acc desc; selected first)

| Variant | LODO H / Acc | AASD H / Acc | ΔAcc vs 53.53 | ΔAcc vs 53.12 | Carta H / Acc |
|---------|--------------|--------------|---------------|---------------|---------------|
| `tip_raw` **←** | 0.7916 / 72.19 | 0.6424 / 53.53 | +0.00 | +0.41 | 0.6460 / 59.71 |
| `sw2_raw` | 0.7914 / 72.09 | 0.6439 / 53.46 | -0.07 | +0.34 | 0.6470 / 59.61 |
| `sw2_isotonic` | 0.7175 / 71.48 | 0.6452 / 53.15 | -0.38 | +0.03 | 0.6601 / 58.97 |
| `sw2_platt` | 0.7271 / 71.46 | 0.6439 / 53.22 | -0.31 | +0.10 | 0.6620 / 59.01 |
| `tip_isotonic` | 0.7172 / 71.31 | 0.6416 / 53.12 | -0.41 | +0.00 | 0.6615 / 59.23 |
| `tip_platt` | 0.7203 / 71.19 | 0.6445 / 53.26 | -0.27 | +0.14 | 0.6621 / 59.26 |

## Targets

| Target | Result |
|---|---|
| AASD Acc vs tip 53.53 / hp30 53.12 | Acc **53.53** (Δ +0.00 / +0.41); H 0.6424 |
| Carta vs linear_hp30 60.54 | Acc **59.71** (Δ -0.83); H 0.6460 |
| Acc>60? | **NO** (AASD Acc 53.53) |
| LODO mean | H 0.7916 Acc 72.19 |

## Verdict

Honest LODO Acc/H selector on train-fold Platt/isotonic of Acc tip SW2 / `hard_abs_swmix|t1` does not create AASD Acc>60. Same qualitative flatness as stem Platt on logits and online-moment on tip streams (~53.46–53.53).

## Do not repeat

Do not rerun train-fold Platt/isotonic on frozen SW2 fuse a=0.5 or `hard_abs_swmix|t1` tip streams at this HMM. Do not re-grid SW2↔swmix soft/hard. Do not confuse with stem `results_calib.md` Platt.

## NEXT IDEA

Calibration of Acc tip decision streams (Platt/isotonic) does not unlock AASD Acc>60 — same collapse as online-moment and domain-emis (all select raw/const tip 53.53). Bottleneck is not train-fold score→label mapping of these streams. Prefer a **qualitatively new** Acc front-end that clears chance and lifts AASD acc1 (speech-path / identity residual into SPK, not another tip post-process). Keep tip AASD Acc 53.53 and `linear_hp30` Carta 60.54 as portfolio leaders.
