# V-REx from IRM stem

- Meaningful change: mean environment BCE plus λ times variance of per-dataset
  risks. This is not IRMv1, Group-DRO, CVaR, or per-dataset upweighting.
- Smoke λ 0.3 and 1.0 were stable; λ 1.0 was used for the full run.
- Train mean passes the aggregate target, but AVGC remains below the per-fold
  target. External transfer regressed, so this technique is closed.
- Diff summary: added `DRO_MODE=vrex`, `VREX_LAMBDA`, isolated run names, and
  uniform environment-risk variance optimization to `stem_ms96_group_dro.py`.
- Raw result: `results/research/stem_ms96_vrex.json`.
- Checkpoints: `models/stem_ms96_vrex/`.
