Repository navigation
Code for DischargeBench: Pre-Deployment Evaluation of LLM Discharge Educators through Persona-Grounded Patient Simulation (arXiv:2609.20827).
v1.0.1 makes failures visible in new runs; no metric definition changes. The paper's results and the PhysioNet simulation records were produced with v1.0.0.
- Simulation Orchestrator: failed judge calls are retried, then recorded as
env_agent_errorinstead of a substituted default verdict;severityrestricted to minor / moderate / severe - Empty or truncated educator replies end the case with an error
- Qwen3 educator replies keep their line breaks
- Resume replaces errored records instead of duplicating them
- SCS: pre-conversation questions removed; failed questions are recorded as errors, not graded incorrect
commission_summary.csvremoved
Full list: "Changes in v1.0.1" in the README.