Resource-Versioned Outcome-Aware Metacontrol for Software Action Systems

Software agents must select admissible actions quickly while deciding whether uncertainty or anticipated consequences justify slower computation. We specify a resource-versioned, outcome-aware metacontroller combining recurrent action preference, bounded lookahead, outcome-sensitive delegation, asynchronous planner proposals, and validation against current resource revisions. On 1,200 core episodes, the proposed method solved 1,160 while a guarded finite-state controller solved 1,200; the proposed method's median decision latency was 18.13 ms versus 0.0219 ms. Four component ablations matched its task success, the fitted deferral gate selected never-defer, and a 24-episode transfer comparison favored XGBoost. The results do not demonstrate an advantage on the evaluated benchmark. We describe the method contract, implementation, results, limits, and experiments needed to test the allocation hypothesis. Code and computational records: https://github.com/fsantibanezleal/CAOS_Neuraxis. License: Apache-2.0 for software; CC BY 4.0 for this preprint.