Runtime Regime Infrastructure: A Systems Architecture for Declared-Regime Verification and Consequential AI

Abstract Modern AI systems possess increasingly sophisticated generation pipelines but little shared computational infrastructure for determining when generated proposals become consequential commitments. Evaluation remains largely implicit—distributed across application logic, organizational review, institutional practice, and post hoc reconstruction—so verification, governance, and dispute resolution often require reconstructing how a decision was made rather than independently recomputing it. This paper presents Runtime Regime Infrastructure (RRI), a derived architecture for the missing computational phase between probabilistic generation and consequential action. Under a declared regime, identity and admissibility are established before evaluation; deterministic evaluation produces typed structural verdicts; policy composes over those verdicts without redefining them; portable decision artifacts preserve the dependencies required to support the result; and independent evaluators recompute the recorded result under a declared replay mode. The architecture follows from a constraint on bounded evaluators: a generated output cannot itself determine the conditions under which it acquires consequential authority. Operational terms such as same, safe, good, and allowed become reproducible only relative to a regime specifying the identity-bearing object, admissible transformations, invariants, evidence boundary, governing authority, and available verdict vocabulary. RRI therefore separates generation from commitment: a proposal becomes actionable only through evaluation under conditions declared independently of that proposal and preserved in an independently recomputable decision artifact. The paper introduces no new formal mathematics. It assembles previously established results from the Identity–Persistence Program—including sufficient regime specification, identity forcing, admissibility, deterministic evaluation, enforcement coding, bounded verification, and decision-artifact obligations—into a single runtime architecture. Each imported result retains its original scope and claim status; the architecture itself is explicitly a DERIVED / ARCHITECTURAL synthesis. The contribution is architectural rather than theorem-level: a layered runtime, typed computational contracts, authority boundaries, refusal semantics, verification model, governance interfaces, and explicit dependency structure for independently reproducible consequential computation under declared regimes.