MAED: Governing the Right to Act in Decision-Capable AI Systems

Large Language Models (LLMs) are increasingly deployed in domains where their outputs implicitly authorize action — financial allocation, risk assessment, operational decisions. Current safety approaches focus on correctness, uncertainty signaling, or alignment with user intent. This paper argues that these approaches fail to address a more fundamental problem: the absence of a governance mechanism that determines whether an AI system is epistemically permitted to act at all. We introduce the Minimum Acceptable Evidence Doctrine (MAED) — a decision-gating framework that governs the right to action rather than the quality of response. MAED enforces a strict separation between analytical capability and decision legitimacy by requiring a minimum threshold of epistemic sufficiency before any action-enabling output is allowed. When this threshold is not met, the system must refuse action, terminate analysis, or downgrade output to a strictly non-actionable form. MAED is presented as an operational extension of the Cognitive Reflexivity Function (CRF), a foundational epistemic constraint that defines conditions under which continued reasoning becomes illegitimate. While CRF specifies when cognition must halt, MAED specifies what minimal, non-deceptive information may be disclosed without violating epistemic integrity. Together, they form a layered governance architecture for decision-capable AI systems. We further describe three enforcement primitives that operationalize MAED in practice: Honesty Boot (pre-analysis epistemic calibration), Right to Refuse (action denial as a first-class system behavior), and Hard Termination (forced cessation of reasoning without follow-up guidance). Using financial decision-making as a high-stakes testbed, we demonstrate how MAED prevents confident but illegitimate outputs that traditional safety and alignment methods routinely permit. This work reframes refusal, silence, and termination not as failures of intelligence, but as necessary features of responsible cognition. MAED represents a shift from optimizing answers to governing action under uncertainty, offering a transferable design pattern for AI systems operating in high-impact environments.