Authorization Architectures for AI-Driven Critical Infrastructure: Runtime Authorization for AI-Assisted Decision Systems in Nuclear and Renewable Energy Environments

Abstract Artificial intelligence is entering critical infrastructure and energy operations at an accelerating rate, assuming roles in forecasting, maintenance optimization, dispatch recommendations, and anomaly interpretation. Existing assurance approaches rely predominantly on monitoring, post hoc audit, and policy-level oversight. In high-consequence environments such as nuclear facilities and renewable energy grids, post hoc oversight is structurally insufficient: detection after the fact does not prevent unsafe action execution. Monitoring creates evidence of what occurred. Runtime authorization creates evidence of what was permitted before occurrence. This paper argues that the principal safety challenge for AI in critical infrastructure is not model accuracy alone, but action eligibility: the determination, at execution time, of whether a proposed action is permissible under the operative constraint regime. The paper proposes a conceptual reference architecture for runtime authorization in AI-driven critical infrastructure. The architecture introduces a non-bypassable authorization layer between AI-generated recommendations and operational actuation, evaluates candidate actions against safety, regulatory, operational, and role-based constraints, and emits a contemporaneous, tamper-evident authorization artifact for every verdict. Application contexts include nuclear facility decision support, renewable generation dispatch, predictive maintenance, and degraded-state emergency operations. Design principles include fail-closed enforcement, separation of recommendation from authorization, explicit constraint evaluation, non-bypassability of the authorization boundary, and authorization-artifact-based auditability. The paper concludes with an implementation research agenda for translating these architectural principles into operational infrastructure. Provenance Note This paper was accepted for presentation at the 8th International Conference on Nuclear and Renewable Energy Resources (NURER 2026), Almaty, Kazakhstan, following peer review in March 2026. It is released here as an open deposit in the FERZ research corpus rather than through the conference proceedings. The peer-review acceptance is reflected in the provenance only; the canonical version of record is this Zenodo deposit. Position in the FERZ Research Corpus This paper applies the architectural arguments developed in earlier corpus entries to the critical infrastructure and energy vertical. The three-verdict enforcement model (ALLOW, DENY, ABSTAIN), the non-bypassability requirement, the authorization artifact concept, and the monitoring-versus-authorization distinction follow the canonical FERZ doctrine established in prior corpus entries on the impossibility of observability-based authorization, the distinction between observability and enforcement, the transition from monitoring to authorization, and execution-time authorization for AI systems. The audit-artifact properties described in Section 4.5 align with the open Four Tests Standard (4TS) for authorization artifacts.