Retrieval-Augmented Reasoning under Constitutional Constraints

Retrieval-augmented systems can expand the information available to a model, but retrieval quality alone does not determine whether a source is current, authorized for access, appropriate for the question, admissible for a particular claim, safe to expose, or faithfully represented in the answer. Retrieval-Augmented Reasoning under Constitutional Constraints (RAR-CC) separates candidate retrieval, access authorization, source characterization, purpose-specific admission, reasoning, claim-evidence binding, review, and external effect. The architecture does not presume that conventional RAG is uniformly ungoverned or that adding receipts produces legitimacy. A receipt can reproduce a biased search; an admitted source can be false; a current source can be irrelevant; a revoked source can remain historically important; and an answer can misstate accurately retrieved evidence. Retrieval replay and answer replay are distinct, and neither establishes correctness. This paper defines an exact retrieval receipt, multi-dimensional evidence artifact, eight evidence postures, context and threat boundaries, multi-policy comparison, claim-evidence map, refusal and uncertainty behavior, governed Hyperpedia/GMB integration, realization funnel, and evaluation program. The goal is not legitimacy optimization but bounded traceability, contestability, and policy-relative evidence use.