This preprint introduces a deterministic execution authority framework designed for AI-mediated and capital-intensive systems operating under irreversibility and uncertainty constraints. The study formalizes a separation between prediction and execution, arguing that probabilistic estimation alone is insufficient for legitimate action in high-impact domains. A minimal governance state vector is defined, incorporating probability, impact, irreversibility, uncertainty, authority adequacy, and time exposure. A log-domain hybrid risk aggregation model is proposed, together with a hybrid authority construct integrating organizational tier validation, cryptographic signature integrity, and a quantitative authority adequacy score. Execution decisions are compiled through a deterministic, fail-closed gate that enforces uncertainty bounds, authority thresholds, and irreversibility-scaled risk constraints. Formal safety invariants are presented, including determinism under canonical input, irreversibility monotonicity, and authority score non-increasing behavior with respect to risk amplification. The framework is compatible with formal specification and model-checking methodologies and structurally aligned with contemporary risk-based AI governance principles. This work does not propose a new predictive model. Instead, it defines an execution governance architecture intended to increase auditability, systemic stability, and regulatory defensibility in AI-driven decision environments.

Version 1.0 (IEEE-style preprint). This manuscript presents a theoretical and governance-focused framework and is not empirically calibrated. The proposed execution authority model is intended as a formal architecture layer for deterministic, fail-closed decision authorization in high-impact AI and capital-intensive environments. The work is released for academic discussion, formal review, and further theoretical and applied extensions, including integration with formal verification toolchains and regulatory compliance infrastructures. No proprietary datasets are used. No commercial product claims are made.