Evaluation artifacts for "A Multi-Agent Large Language Model System for Epidemiological Intelligence: Design, Deployment, and a Controlled Evaluation of When Orchestration Helps"

Reproducibility bundle for the JAMIA Open manuscript "A Multi-Agent Large Language Model System for Epidemiological Intelligence." Contains the 50-question benchmark (7 categories), the runner and LLM-as-judge scripts, and the raw per-question results and judge scores for every configuration (multi-agent, single-agent, and three ablations), so the reported statistics are fully reproducible. Agents ran on Claude Opus 4.8; the judge used GPT-4o at temperature 0. The EpiGIS Pro platform source is private and not included; it is described in the cited platform paper (Int J Health Geogr, 2026, doi:10.1186/s12942-026-00479-1). No credentials or endpoints are included.