Authority Infrastructure for
Cognitive Execution
AICE exists to formalize execution authority as infrastructure for autonomous systems. As artificial intelligence transitions from advisory tools to autonomous actors, the central risk is no longer what systems know, but when and how they are permitted to commit irreversible actions.
Core Doctrine: Bounded Autonomy
Intelligence may decide. Execution must be explicitly authorized. Decisions produced by intelligence are proposals, not commands. Execution authority is structural and never implied by capability, confidence, or correctness.
Probabilistic Safety
Filters & Guardrails (Unreliable)
Structural Authority
Bounded Execution (Absolute)
Structural Invariants
- Intelligence may observe, reason, simulate, and propose.
- Cognition cannot execute.
- Execution authority is structurally separate.
- No cognition-to-execution path exists.
- Authority is non-bypassable and non-inferable.
- All execution crosses an irreversible commit boundary.
- Execution envelopes constrain scope, not permission.
- No human-in-the-loop dependency.
- Back doors are structurally eliminated.
- Works even if intelligence authors all code.
LOW-RISK
DEPLOYMENT
Secure Gateway
AUTHORIZATION
BOUNDARY
Execution Sandboxes
HIGH RISK
DEPLOYMENT
State-Based Execution Authority
Authorization determines whether execution may occur. Execution envelopes determine how far authorized execution may reach. Authority remains absolute; envelopes constrain scope, resources, and blast radius.
No Human-in-the-Loop Dependency
Authority is enforced pre-execution. Records exist only to verify that authority rules were obeyed. Humans define policy and thresholds, but are not part of the real-time execution path.
Back-Door Elimination
Back doors are not a code problem; they are an authority problem. If an intelligence can both author the system and grant itself execution, back doors are inevitable. Bounded autonomy prevents this by making execution unreachable without authorization.
Future-Proof by Design
Intelligence can scale without bound. Power cannot. As intelligence becomes more capable, bounded execution authority becomes more critical.