Mode-Bounded Intelligence: Structural Constraints on Action in High-Capability Systems

Approaches to AI safety and organisational reliability focus primarily on aligning intent, values or objectives. Many failures of intelligent systems arise not from misalignment or malice but from the ability to act across incompatible operating contexts without explicit constraint. This paper introduces mode-bounded intelligence: an architectural approach in which intelligent systems are constrained by explicit operating modes governing authority, commitment formation and reversibility of action. As capability increases, discretion must decrease and procedure must increase — not as human bureaucracy but as machine-enforced structural constraint. Modes are properties of system state rather than of agent identity, defined by their stance toward uncertainty, scope of authority, permissible commitments and tolerance for failure. We specify a minimal mode set, state the transition rules as invariants over system state, and treat commitments and authority as first-class objects external to intelligence itself. Doing so prevents a class of failure in which exploration, execution, adjudication and architectural change collapse into one undifferentiated process. Two limits are stated rather than assumed. The invariants bind only where an interlock converts a verdict into a constraint on action, which the one instrumented deployment of this pattern establishes empirically; and the minimal mode set is minimal for the human--machine institution rather than for the machine, since modes not instantiated in the system are borne by people. This is a design constraint, not a doctrine.