Distinction, Execution, and Adjudication

Abstract

This paper introduces DEA (Distinction, Execution, and Adjudication), a governance architecture for LLM-assisted scientific workflows. DEA separates the authority to define a scientific transition, the capability to execute it, and the authority to adjudicate its result. The framework combines capability separation, evidence boundaries, explicit authorization artifacts, and the No Authorized Path (NAP) state to reduce self-confirmation and unauthorized transitions. The work connects agentic AI governance with scientific workflow management, reproducibility, security principles, and epistemic governance. Rather than claiming epistemic certainty, DEA aims to improve the traceability, containment, and independent assessment of scientific work performed with AI systems. The paper presents the conceptual architecture and a prospective evaluation design comparing monolithic, prompt-separated, and capability-separated workflows; empirical validation is reserved for a subsequent pilot study.