# results_mixup_stem_avgc

Input/label **Mixup** fine-tune of `stem_ms96` from **focal-IRM**. Sample `lam ~ Beta(α,α)` (clipped ≥0.5), mix ms windows + soft labels; keep recipe BCE + post-switch weights + TV. Architecture unchanged (ms, w96).
Init `models/stem_ms96_focal_avgc/{held-*,all5}.pt`; causal `-inf` mask restored. EMA 0.995.
Smoke α∈{0.2,0.4} (40 steps, held-AVGC) → **α=0.4** (frame-acc 0.948).
Full 800 steps, batch 512. HMM 0.994/0.8 + hp30. Also score **fuse w=0.5** with frozen `stem_synth_switch`.

**Distinct from** MixStyle-on-vlaai, AVGC λ upweight/curriculum, focal γ re-grid, soft-H, fuse-conf-resid, adapt-readout.

Code: `stem_ms96_mixup_avgc.py`. Checkpoints: `models/stem_ms96_mixup_avgc/`.
JSON: `results/research/stem_ms96_mixup_avgc.json`.
Log: `experiments/stem_ms96_mixup_avgc.log` / `_console.log`.

## LODO (preset)

| Held-out | mixup H | Acc % | fuse_w0.5 H/Acc | tip focal↔synth | hp30 H | hp30 Acc % |
|---|---:|---:|---|---|---:|---:|
| AVGC | 0.8162 | 74.70 | **0.8213 / 75.29** | **0.8226 / 75.54** | 0.7839 | 71.39 |
| DTU | 0.9097 | 88.63 | 0.9199 / 90.23 | — | 0.8445 | 73.61 |
| KUL | 0.8782 | 81.94 | 0.8878 / 82.96 | — | 0.8402 | 73.91 |
| UHD | 0.8586 | 79.71 | 0.8573 / 80.48 | 0.8587 | 0.8221 | 74.06 |
| USTC | 0.9146 | 85.58 | 0.9214 / 86.62 | — | 0.8422 | 72.82 |
| **mean** | **0.8755** | **82.11** | **0.8815 / 83.12** | **0.8821 / 83.36** | | |

## External (preset)

| Set | mixup H/Acc | fuse_w0.5 H/Acc | linear_hp30 | tip focal↔synth fuse |
|---|---|---|---|---|
| AASD | 0.6185 / 50.84 | 0.6215 / 50.96 | 0.6422/53.12 | 0.6208/51.06 |
| Carta2026 | 0.5465 / 48.80 | 0.5565 / 49.56 | 0.6520/60.54 | 0.5496/49.14 |

## Targets

| Target | Result |
|---|---|
| AVGC ≥0.85 / Acc≥78 | **miss** — pure mixup 0.8162/74.70; fuse_w0.5 **0.8213/75.29** (Δtip **−0.0013 H / −0.25 Acc** vs 0.8226/75.54) |
| Beat tip fuse | **no** — fuse mean 0.8815/83.12 slightly below tip 0.8821/83.36 |
| AASD Acc>60 | miss (~51) |

## Notes

- Smoke preferred α=0.4 over 0.2 (frame-acc 0.948 vs 0.946).
- Pure mixup AVGC below tip; fuse w=0.5 recovers most of the tip but does not beat it.
- Input Mixup on ms stem saturates like other stem regularizers (stochdepth / SAM / soft-H) near the focal↔synth envelope.
- AASD Acc frozen ~51.

## NEXT IDEA

Stop stem Mixup / BCE-family re-tunes from tip inits — AVGC tip envelope is flat. Prefer **Acc-only** front-end that does not touch LODO tip (e.g. short-hold / noise-aug on published SW2 if still open), or a **qualitatively new stem evidence stream** (not Mixup/focal/λ) before another fuse w-grid.
