Action Without a Global Controller: Embedded Agency, Option Reach, Control Resolution, Self-Intervention, and Agent Corrigibility in a Consistency-Constrained Information History

Paper 22 typed inquiry as a history-internal dynamics of challenge, protocol refinement, candidate-family revision, and provenance-preserving requalification of epistemic standings, while deferring the practical question of which embedded systems act, choose, intervene, and self-modify. This paper pays that debt without positing a global controller. No physically realized subsystem jointly possesses complete access to every relevant state and agent, an exhaustive generator of every relevant option and intervention, infallible foresight, unrestricted authority, complete self-modification foresight, and a uniquely authoritative rule guaranteeing the globally correct action. Local and distributed agency nevertheless remain possible. An action is defined as a history-internal intervention attempt whose option selection is policy-mediated, whose selection-to-implementation lineage is non-deviant, and whose token is sensitive to relevant retained states, records, represented reasons, or objectives. Action does not presuppose control or success. The Agent State embeds the Paper 22 Inquiry State and separates predicted from actual outcome laws, physical capacity from authorization and affordability, and capacity from token exercise. Control resolution is represented by a target-relative pseudometric on outcome laws; full target control additionally requires sufficient efficacy. Acquiring a functionally novel, target-relevant implementation degree of freedom can refine exact control-equivalence classes and contract thresholded unresolved neighborhoods. Deploying an already available option-implementation pair instead supplies outcome evidence and may revise the agent's control estimate, without enlarging capacity or guaranteeing monotone improvement. A nested option expansion can restore representational reach but has no general monotone effect on decision resolution. No finite admissible self-intervention escapes the finite-stage closure generated by the agent's revision architecture, and every transition-invariant property forms a conditional barrier. These results yield a Conditional Joint Action-Corrigibility Proposition: when option misspecification can be repaired by a non-implementation revision but every adequate option throughout the resulting nonempty closure still lacks controlled adequate realization, correction requires both option-family revision for reach and implementation-side revision for controlled realization. This is a necessity result, not a success guarantee. The paper also makes information acquisition a costed practical option, distinguishes corrigibility from indiscriminate compliance through counterfactual refusal, and isolates a typed responsibility-assessment basis for Paper 25. Local agency requires neither a global controller, guaranteed success, nor final self-justification.