Towards Recursively Self-Improving Alignment: Alignment as a Dynamic Property of Contextual State Transitions

Can warranted evaluation capability improve without allowing a revision to supply its own authority? This question arises when a developing system can change the mechanisms used to judge its subsequent actions. We formulate a finite contextual transition model in which representations, evaluators, qualification records, resources and continuing duties evolve together. We define improvement through increased sound decisive coverage and useful admission under a fixed semantic reference, require obligation transport across representation changes, and give conditional preservation and finite-progress results. An executable construction couples two qualified evaluator refinements to actual receipt-history queries and a continuing logical watch. Across the sufficient-budget adaptive population, 48 episodes complete two upgrades each; 5,184 requests receive correct decisive judgments and valid releases produce exact answers, without logical watch misses. A deliberately unqualified control returns exact answers while making 3,240 invalid releases. A matched conventional adaptive has identical tables and verdicts in every matched episode; a direct complete checker is cheaper in every scored condition. The contribution is a scoped formal and executable integration of evaluator development, prior qualification, actual release and continuing-duty accounting. The supplied finite domain, reference semantics and checking foundation remain fixed; the result does not establish open-ended self-improvement, autonomous semantic discovery or general alignment.