Toward an Audit-Ready, Constraint-Based Architecture for Oncology Clinical Decision Support with Large Language Models

This preprint describes a constraint-based, audit-ready architectural framework for integrating large language models into oncology clinical decision support. The work focuses on epistemic governance rather than clinical performance, proposing a system design that externalizes context handling, evidence authority, and verification outside the language model. The architecture combines a Context Utility Layer (CUL) and a Truth-Checker Layer (TCL) to enforce jurisdiction awareness, guideline conditionality, traceability, and fail-closed behavior in high-stakes clinical settings. The manuscript is a design specification and implementation walkthrough. It does not report clinical outcomes, does not automate treatment decisions, and does not propose autonomous AI behavior. Its scope is limited to auditability, regulatory alignment, and safe system-level integration of LLMs in oncology decision support.