CARA: Constrained Artificial Reasoning Architecture

This paper introduces CARA (Constrained Artificial Reasoning Architecture), a protocol-level governance framework for reasoning systems, including AI systems, that treats epistemic authority as an explicitly managed state rather than an emergent property of model behavior. CARA identifies a class of failures termed ‘authority laundering,’ where claims or artifacts silently gain legitimacy without any auditable authority‑granting event, and proposes a small set of primitives—trace‑first execution, phase discipline, explicit admission events, producer–admitter separation, and fail‑closed transitions—to prevent unauthorized epistemic state changes. The paper specifies minimal entities and invariants, along with diagnostic canaries that distinguish governed from ungoverned systems, while intentionally omitting implementation details so the focus remains on information discipline rather than a particular deployment.