Beyond Optimization: Entropy-Regulated Artificial Cognition and the Problem of Long-Horizon Stability A Theoretical Framework for Organizational Regulation in Recursive Intelligence

This preprint introduces Entropy-Regulated Artificial Cognition (ERAC), a theoretical framework that distinguishes capability-generating dynamics from organization-preserving dynamics in recursive artificial intelligence. The paper defines semantic collapse as a long-horizon organizational process in which an artificial cognitive system retains local task competence while its available semantic trajectories become increasingly concentrated, repetitive, and difficult to revise. ERAC argues that sustainable artificial cognition requires regulation within an adaptive entropy region rather than pure external optimization. In particular, the framework proposes that outer regulation alone is insufficient for long-horizon stability and that dual regulation (inner and outer) must be established from the outset. Empirical illustrations are provided using large-scale recursive generation experiments with 32B-class models, 16,384-token contexts, and n=20 runs. The results indicate that the dual-hybrid condition produces higher lexical diversity, lower repetition, and reduced semantic drift compared with single-regulation baselines, while exhibiting expected trade-offs in local coherence and prompt alignment. This version is a complete draft intended for early dissemination and feedback. Code and full experimental artifacts are not included. A substantially revised and condensed version is planned for journal submission.