From Knowledge to Control A Decision Engineering Architecture for Admissible AI System A unified framework integrating epistemic, predictive, and governance layers through decision admissibility

This working paper introduces a unified architectural framework for artificial intelligence systems that integrates epistemic representation, predictive modeling, decision processes, and execution control into a single coherent structure. While recent advances in AI have significantly improved knowledge representation and predictive performance, current systems remain structurally incomplete. They are optimized to estimate what is likely to happen, but lack a formal mechanism to determine what is allowed to happen. This gap results in a fundamental limitation: high-quality predictions and decisions can still lead to inadmissible or harmful outcomes when executed in dynamic environments. To address this limitation, the paper proposes Decision Engineering Science™ (DES) as a foundational layer for transforming intelligence into controlled action. Within this framework, a decision system is defined as a mapping from state and knowledge to a constrained space of admissible actions. The architecture introduces a formal decision layer, responsible for defining admissibility through constraints, objectives, and governance structures. A key contribution of the paper is the concept of commit-time admissibility, operationalized through the execution boundary (CARE). This mechanism ensures that decisions remain valid at the moment of execution, accounting for changes in context, constraints, and risk conditions. Only actions that satisfy admissibility at commit time are allowed to be realized, establishing a direct link between decision logic and real-world outcomes. The proposed framework unifies six layers: Human layer (objectives, values, decision ownership) Epistemic layer (knowledge representation and uncertainty) Predictive layer (modeling of possible futures) Decision layer (admissibility and action selection) Execution layer (realization of state transitions) Governance layer (constraints, compliance, and oversight) By embedding admissibility into system architecture, the framework shifts AI design from prediction optimization to outcome control. This enables systems that are not only intelligent, but also structurally aligned with regulatory requirements, including the EU AI Act, and capable of enforcing constraints in real time. The paper contributes to the emerging field of decision-centric AI by formalizing the relationship between knowledge, prediction, and action, and by introducing a control-oriented perspective on system design. It provides both theoretical foundations and architectural principles for building AI systems that ensure only admissible states become real.