andyjoecn/TrustMed-RL-eval-baselines

TrustMed-RL test-set rollouts of external baselines

Outputs of Eval/code/baselines/run_arm.py (frozen consultation environment, crops-only evidence, gpt-4.1-mini patient) on the frozen 2,500-case test split, one directory per arm and pass: <arm_id>/pass<N>/clean/{trajectories.jsonl, summary.json, arm.json, run_manifest.json, driver.log, state/usage.jsonl}. No judge applied; score with baselines/cluster/judge_and_score.sh. Mismatch segments (mm_*) are pending the integrity-gate donor bank.