A confident model isn’t a safe one. In my open-source (MIT) LangGraph underwriting agent, a risk ceiling in plain code overrides every model, so prompt injection can’t get a risky loan approved. Code: github.com/shadrach-tayo/underwriting-agent