From Minimal Model Systems to World Models: Boundary-Relative Distinctions for Factoring Generative Models

The good regulator theorem holds that a successful regulator must embody a model of the system it regulates, and recent work suggests that even minimal regulators can carry informative world models. We take up a prior structural question: which distinctions should organize the channels through which a model's content arrives? We compare four modeling starting points, Turing's unorganized machines, Ashby's homeostat, autopoiesis, and active inference, for the distinct elements and constraints each makes salient. From this comparison we derive two candidate sets of primitives. Causal Standpoint Routes (CSR) classify a causal relation by whether its source and target lie inside or outside a declared system boundary, and by whether the relation is immediate or mediated relative to declared operations and interfaces, which gives eight route classes. Allostatic Kinds distinguish four directions of coupling between a regulator and its environment and relate each to the essential variables and viable bounds that give it regulatory significance. Unorganized machines offer a particularly direct minimal model for CSR, and the homeostat for Allostatic Kinds. We outline how both could inform the factorization of generative models in active inference, for instance by separating external sources, intermediaries, and sensory evidence, and we propose an empirical program testing whether CSR-guided training or steering improves held-out causal reasoning in language models against matched controls. The correspondences among the four frameworks remain open, and we present the primitives as testable hypotheses.