Pre-Reasoning Gating Clarifying the Boundary Between Provisional Computation and Authorized Commitment

Large language model systems are increasingly deployed in contexts where outputs can trigger irreversible consequences: financial transactions, regulatory submissions, infrastructure changes, and autonomous operations. Current architectures typically permit commitment-capable operation by default, relying on post-generation filtering to prevent harm. We argue this default is structurally misaligned with high-stakes deployment requirements. This paper defines Pre-Reasoning Gating (PRG) as an architectural boundary requiring explicit authorization before a system may enter a mode permitting commitment-class actions—those whose effects are expensive to reverse or predictably induce real-world state change. PRG is not a novel mechanism; it names a familiar primitive from safety engineering and access control, adapted to inference systems. The contribution is definitional: we formalize the commitment set as actions above a reversibility threshold, distinguish commitment-capable modes from advisory modes, include induced commitments (outputs that reliably trigger downstream execution without independent gates), and provide crisp exclusions separating PRG from monitoring, throttling, rollback, and post-hoc filtering. Through concrete failure cases in tool-use systems, we demonstrate why the distinction between mode-entry authorization and output filtering matters in practice. PRG applies narrowly—only where incorrect action costs substantially exceed delayed action and reversal is expensive—but within that scope, it isolates a control boundary frequently collapsed in current system designs.