GRPO run: grpo-20260426-062846

train: torch_dtype not dtype

README: judging-criteria mapping + reproduce + per-task evidence pointers

notebook: add executed cell outputs (smoke test + job submission). Chart cells fill on re-execute after job completes.

train: drop Unsloth (CUDA-init flake), bypass is_available cache, default to Qwen/Qwen2.5-1.5B-Instruct

train: CUDA-warmup retry + cuda-gated unsloth import

train: drop server/* import — inline verb table + dense costs (job ran without openenv-core)

train: fix WS protocol — use data wrapper + flat message field

claims-env-pro v2.0.0 — initial port (12 verbs / 5 tasks / 18 obs / 6-comp reward / Plaid)

initial commit