Sovereign Acceptance Runtime: Verifiable Human Ratification and Evidentiary Audit Trails for Autonomous AI Agents under the EU AI Act

Autonomous AI agents are rapidly evolving from conversational assistants into software systems capable of invoking external services, modifying persistent state, and executing transactions with real-world consequences. While contemporary agent frameworks frequently incorporate human-in-the-loop (HITL) approval workflows, these mechanisms typically provide operational control without generating portable, cryptographically verifiable evidence of who authorized an action, which execution payload was reviewed, and whether the executed action remained identical to the approved one. The regulatory significance of this limitation has increased with Regulation (EU) 2024/1689 (the EU AI Act). Under Article 113 as amended by Regulation (EU) 2026/1744 (the Digital Omnibus on AI), high-risk AI systems become subject to mandatory requirements for automatic event logging, technically implemented human oversight, and operational log retention from 2 December 2027 (systems listed in Annex III) and 2 August 2028 (systems covered by Annex I Union harmonisation legislation). Although these obligations establish regulatory objectives, neither the Regulation nor currently available technical standards specify how authorization records should acquire independent evidentiary value or cryptographic resistance to undetected modification. This creates an architectural gap between regulatory requirements and implementation guidance. This paper introduces the Sovereign Acceptance Runtime (SAR), an execution architecture that models human authorization as a cryptographically verifiable ratification bound to the exact execution payload. Concurrent IETF Internet-Drafts already define pre-execution authorization records and human-signed action approvals; SAR's contribution is their composition into a single profile. Building upon the IETF Supply Chain Integrity, Transparency, and Trust (SCITT) architecture (RFC 9943) and CBOR Object Signing and Encryption (COSE) signed statements, SAR represents authorization as a machine-enforceable Registration Policy evaluated before execution. The runtime adopts a fail-closed execution model in which actions proceed only when accompanied by verifiable authorization evidence whose payload binding remains valid at the moment of execution. The paper develops a threat model and six architectural requirements for verifiable human ratification, describes a reference architecture implementing these requirements, positions the work against concurrent IETF drafts on agent authorization, relates the proposed mechanisms to selected provisions of the EU AI Act, evaluates a reference implementation using COSE_Sign1 and deterministic CBOR against an implementation-independent conformance suite whose adequacy is checked by mutation analysis, and outlines a potential path toward interoperability within the IETF SCITT ecosystem.