PRANA
PRANA is an environment for training and evaluating AI agents on clinical administrative workflows.
Core trick
Synthetic transplant-candidate records carry expiry dates, so agents must re-query stale labs and pass automated validators on regulatory reports.
Description
PRANA is an environment for training and evaluating AI agents on clinical administrative workflows. Tasks such as patient triage and transplant management involve long-horizon reasoning, fragmented data sources, and strict regulatory validation.
PRANA simulates these workflows through stochastic patient data and multiple scenario trajectories, allowing agents to query records, reconcile missing or stale information, and complete regulatory reporting tasks.
The environment supports scalable task generation via synthetic patient instances and structured validators that automatically evaluate outcomes.
To measure agent performance, I introduce a τ² (Tau-Squared) benchmark tailored to clinical workflows, designed to test long-horizon reasoning, temporal validity of medical data, and efficient information retrieval. Early results show that even frontier models struggle with the hardest τ² scenarios.
README excerpt
--- title: PRANA-Env Environment Server emoji: 🏥 colorFrom: purple colorTo: indigo sdk: docker pinned: false app_port: 7860 base_path: /web tags: - openenv - reinforcement-learning - clinical --- # PRANA-Env **Policy Reinforced Administrative Navigation Agent** — an RL environment for clinical administrative workflow agents, applied to kidney transplant candidate management. > [!IMPORTANT] > **All data in this repository is entirely synthetic.** Every patient record is procedurally generated by `data/generate_patients.py` and contains no real patient data or protected health information (PHI). Any resemblance to real individuals is coincidental. Lifelong treatments such as organ transplants involve inherently long-horizon administrative workflows: labs expire, records fragment across systems, regulatory deadlines are unforgiving, and clinical anomalies must be caught before filing. PRANA-Env models this complexity as a reinforcement learning problem, providing a stochastic, multi-scenario environment where an agent learns to navigate real-world clinical constraints rather than memorize fixed trajectories. The agent must query fragmented clinical datastores, reason about temporal validity of lab results, detect anomalous measurements per OPTN policy, and file a complete KARS-compliant SRTR report — receiving reward from a deterministic validator that mirrors actual regulatory requirements. A companion **tau2 benchmark** measures agent performance on temporal reason