Deterministic State-Transition Control for Recursive AI R&D: Separating Proposal, Verification, Authorization, and Commit in Self-Modifying Research Loops

Recursive AI R&D creates a distinct systems-control problem when the artifact being optimized can include the agent, research harness, evaluator, training procedure, policy surface, or deployment system that determines which future changes are accepted. This paper develops a deterministic state-transition control architecture for recursive and self-modifying AI research loops. It separates four functions that are frequently collapsed in automated development systems: proposal, verification, authorization, and commit. An AI-generated proposal may produce a candidate change. Verification may establish whether specified technical predicates are supported by evidence. Authorization determines whether the requested operation is permitted for the relevant subject, object, scope, generation, time, and context. A separate commit boundary alone performs the protected state mutation. No earlier stage is permitted to inherit the authority of a later stage. The framework models recursive AI R&D as a protected state-transition system rather than as an ordinary continuous-integration pipeline. It introduces operation-specific technical determinacy, protected-core separation, generation binding, non-inheritance of authority across operations, evidence-lineage preservation, monotonic revocation, rollback constraints, recovery controls, and fresh readmission requirements. The architecture does not attempt to make AI behavior deterministic. Instead, it makes privileged state-transition semantics explicit, independently testable, and enforceable even when the surrounding research process is highly autonomous, adaptive, or recursively self-modifying. The paper further develops a formal state model, authority and transition invariants, multi-agent and cross-provider considerations, failure and race-condition analysis, negative testing requirements, minimum conformance properties, and implementation-neutral standardization criteria. Central non-equivalences include: proposal is not verification; verification is not authorization; authorization is not commit; credential validity is not operation authority; benchmark success is not deployment authority; rollback is not authority restoration; recovery is not automatic readmission. This publication is intentionally limited to public research-level abstractions. It does not disclose unpublished patent claim language, confidential claim charts, private source locators, provider-specific production parameters, non-public test vectors, or other confidential implementation details. Structural Paper Series - Paper 03 Final Publication Edition v2.1 Research Program on Deterministic Infrastructure and Human-Centered AI Coordination Transition Intelligence Institute, Switzerland Institutional Establishment in Preparation Foundational Research Signature: The Second Waters