Built on LangGraph and Gemini, the harness is the orchestration loop that turns a clinician question into a grounded answer: it calls the risk model and retrieval through our MCP server, checks every response against deterministic guardrails, and narrates the result in plain language, with citations vs. hallucinations.