This paper formalizes refusal as an architectural invariant in bureaucratic automation: a system-level constraint requiring that automated or semi-automated administrative systems become structurally incapable of issuing decisions once defined dignity, due process, or epistemic sufficiency thresholds are crossed. Unlike policy safeguards, ethical guidelines, or human-in-the-loop oversight, refusal is specified here as a reachability constraint that operates independently of procedural legitimacy, institutional mandate, or optimization pressure. The paper distinguishes refusal from discretion, escalation, and post-hoc review, and demonstrates why systems that retain the capacity to act under prohibited conditions cannot be considered safe, trustworthy, or governable at scale. By framing refusal as a falsifiable, testable invariant rather than a normative preference, this work establishes prior architectural ground for the design, evaluation, and certification of bureaucratic decision systems. It contributes a foundational constraint for automated governance and positions refusal as a necessary condition for legitimate delegation of authority to machines.
Deusdedit Ruhangariyo (Tue,) studied this question.