The Contextual Imperative: A Utility-Based Framework for Forensic LLM Risk Assessment

Description:As Large Language Models (LLMs) transition from generative assistants to decision-support tools in criminal justice, traditional metrics like AUC-ROC are failing to capture the ethical costs of error. This paper introduces the "Contextual Imperative," a framework that evaluates forensic AI utility as a function of outcome prevalence and the normative "Cost" ratio of the legal environment.Adhering to the TRIPOD-LLM reporting guidelines, we apply Decision Curve Analysis (DCA) to define the "Moral Equilibrium"—the specific probability threshold ($p_t$) where the harm of unnecessary detention equals the benefit of violence prevention. Our results reveal a critical "Prevalence Trap": in general population settings (5% prevalence), even high-performing models yield a Positive Predictive Value (PPV) of less than 30%. Furthermore, we propose a mandatory "Right to Reason" for forensic AI, validated by parallel safety architectures in physical AI (NVIDIA Alpamayo), requiring that any deprivation of liberty be justified by a traceable, verified linguistic rationale.