Large language models (LLM) are being used in educational settings to provide guidance, feedback, and policy-related judgments. However, these systems produce inconsistent responses to similar cases, raising concerns about fairness, transparency, and trust. Rather than relying on probabilistic model output, our approach represents cases in terms of salient features and applies principles derived from institutional academic policies to guide decisions. The LLM remains responsible for interpreting cases, while a separate layer using principles ensures similar cases are treated consistently. This approach offers a middle ground between unrestricted generative AI and rule-based systems. By making the basis for decisions explicit and inspectable, it provides greater transparency, supports accountability,and allows policies to be operationalized in a consistent form. We illustrate the approach using examples drawn from our college policies concerning student use of AI and discuss the implications for responsible deployment within the liberal arts.
