CARA: Constrained Artificial Reasoning Architecture

A Protocol-Level Framework for Governing Epistemic State in Reasoning Systems CARA introduces a protocol-level governance framework for reasoning systems—particularly AI systems—that treats epistemic authority as an explicitly managed state, not an emergent byproduct of model behavior or repeated use. Rather than optimizing reasoning performance, CARA governs when reasoning outputs are allowed to count as authorized, persistent, or reusable state. The framework identifies a class of systemic failures termed authority laundering, in which claims, artifacts, or intermediate results silently acquire legitimacy without any auditable authority-granting event. To prevent this, CARA defines a minimal set of governance primitives—trace-first execution, phase discipline, explicit admission events, producer–admitter separation, and fail-closed transitions—that regulate epistemic state transitions independently of how reasoning is performed. CARA intentionally avoids prescribing implementations, algorithms, or model architectures. Its concern is informational discipline: making epistemic promotion explicit, reviewable, and reversible. The paper specifies invariants and diagnostic canaries that distinguish governed from ungoverned reasoning systems, while leaving deployment choices open. CARA is designed to operate beneath application logic and above raw computation, providing a substrate for systems that must reason under uncertainty without allowing partial evidence, success, or repetition to silently become authority.