# results_stem_sw2_ms

Fine-tune `stem_ms96` from **SWA** with published **SW2 fuse** evidence (a=0.5, DEC_OFFSET=3) concatenated onto 24-d ms features → **26-d**. Zero-init new input columns. EMA 0.995. HMM **0.994/0.8**.

Code: `stem_ms96_sw2_ms.py`. Ckpts: `models/stem_ms96_sw2_ms/`. JSON: `results/research/stem_ms96_sw2_ms.json`.

## LODO (preset)

| Held-out | H | Acc % | hp30 H | hp30 Acc |
|---|---:|---:|---:|---:|
| AVGC | 0.8054 | 75.29 | 0.7762 | 71.06 |
| DTU | 0.9007 | 89.09 | 0.8265 | 70.84 |
| KUL | 0.8684 | 81.84 | 0.8437 | 74.77 |
| UHD | 0.8517 | 80.05 | 0.8096 | 73.31 |
| USTC | 0.9171 | 86.36 | 0.8298 | 70.93 |
| **mean** | **0.8686** | **82.53** | | |

## External (preset)

| Set | sw2_ms H/Acc | hp30 H/Acc | vs linear_hp30 53.12 / SW2 53.46 |
|---|---|---|---|
| AASD | 0.6197/**50.87** | 0.6016/51.11 | Acc 50.87 (Δ -2.25 vs hp30; Δ -2.59 vs SW2) |
| Carta2026 | 0.5433/**47.93** | 0.5613/49.26 | Acc 47.93 |

## Targets

| Target | Result |
|---|---|
| LODO mean | **0.8686/82.53** |
| AVGC ≥0.85/78 | 0.8054/75.29 |
| UHD ≥0.85 | 0.8517 |
| AASD Acc>60 | **50.87** (vs 53.12 / SW2 53.46); H 0.6197 |
| Carta | 0.5433/47.93 |

## Verdict

No open target moved. SW2-as-ms-channel fine-tune does not import SW2's external Acc into the stem stream.

## NEXT IDEA

**xcorr-matched-clf** already abandoned (chance @50 steps). Prefer a new Acc-path front-end that clears chance (or a non-stem evidence transform distinct from SW2/hp30 banks). Do not re-grid SW2↔stem channel weights or reopen fuse↔hp30 / SW2 bank expands.
