Clinical lab interpretation is often a process problem, not a one‑shot recall problem: the same numeric result can be harmless or critical depending on pregnancy, medications, prior baselines, or therapeutic targets. This environment reframes clinical evaluation as a tool‑calling, evidence‑gathering task so models are scored on the medical process they follow rather than on a free‑text diagnosis alone.
What Sets It Apart
- Process‑first grading: rewards are dense and step‑level (demographics lookup, medication checks, reference‑range queries, historical deltas, critical flags) so models must demonstrate a defensible workflow, not just surface recall. This makes error analysis actionable for safety testing and supervision learning.
- Realistic LIMS surface: an in‑memory SQLite schema with relational tables (patients, meds, lab orders/results, previous results, critical alerts, pending cases) forces agents to navigate an interface similar to production clinical systems, including distractor cases and context‑sensitive reference ranges.
- Reproducible runtime and tooling: OpenEnv/FastAPI server, deterministic graders, bounded episode lengths, and optional trajectory export make it suitable for offline RL, policy comparison, prompt engineering, and generating supervised training traces.
Who it’s for and tradeoffs
Great fit if you are evaluating or training agents to perform structured tool use and clinical process reasoning, comparing tool‑calling policies, or generating trajectory datasets for offline RL in a medically grounded domain. Look elsewhere if you need real patient data, broad pathology coverage beyond the provided 8 scenarios, or a certified clinical decision tool—this is a synthetic, research‑oriented benchmark and explicitly not for patient care.
