Governing the Transition from Decision to Learning: A Target Architecture for Warrant-Preserving Adaptive Systems

This paper develops a target architecture for governing the transition of institutional governance determinations into learning inputs for adaptive AI systems. A governance determination may be warranted for its original purpose without being warranted for subsequent use as a training label, reward signal, evaluation target, or other learning input. Accurate provenance, faithful encoding, and a recorded positive decision do not automatically establish authority for this change of function. The proposed architecture treats the transition as a claim-indexed governance object and distinguishes four logical functions: obligation discovery (D), warrant justification (J), governed status (S), and execution authorisation (X). It specifies warrant through evidence, normative authority, and inferential binding (E/N/B), while preserving the distinction between supported positive findings (PASS), unresolved inquiries (OPEN), and supported negative findings (NO). The paper develops analytical countermodels, examines temporal reassessment and descendant standing, and presents P115, a conditional admission-safety invariant. Under a declared state-transition contract, the formal argument establishes preservation of a guarded-admission invariant across reachable states. The result does not independently establish substantive normative warrant, implementation correctness, or correct downstream learning. These remain separate obligations. The contribution is a formally specified research target for warrant-preserving governance-derived learning, making explicit the claim-specific evidential, normative, and inferential obligations that must be examined when a governance determination is proposed for a new learning function, without claiming that those obligations have been discharged.