The EU AI Act establishes a comprehensive risk-based regulatory framework governing AI systems, including obligations relating to risk management, technical documentation, logging, human oversight, and post-market monitoring. For general-purpose AI models (GPAI), additional obligations are supported by European Commission guidance and the GPAI Code of Practice. However, the Act does not prescribe a specific runtime architecture for verifying that authority and admissibility remain valid at the exact moment an AI-enabled system executes an externally consequential action. This paper introduces ARETABA as an execution-time authority and admissibility control architecture for execution-bearing agentic systems. It addresses what is defined here as the execution-time authority gap: the divergence between authority validated at approval and authority that actually exists at the moment of execution. ARETABA is not a legal framework and does not replace the EU AI Act. It is an implementation and assurance architecture designed to support providers, deployers, auditors, and procuring bodies in operationalising existing regulatory obligations. ARETABA operates within a broader architectural stack in which ISDAIRE defines admissibility ex ante, ARETABA constructs and validates runtime authority, and OTANIS enforces authority at execution boundaries.At the governed commit boundary, not only authority but all required execution-time admissibility predicates must hold. If any required predicate fails, is stale, indeterminate, or cannot be resolved within the declared execution constraints, commit must fail closed through refusal or governed escalation.
Masayuki Otani (Wed,) studied this question.