World Models Without Worlds II: Why Learned Models Do Not Refer, and Why Guardrails and Validation Cannot Close the Gap

Since 2025 the phrase "world model" has come to name at least six architectures: latent predictors in the JEPA line, generative video simulators, explicit spatial generators, model-based agents, evaluation simulators for autonomous driving, and hybrids in which a language model directs a generator. This paper asks one question of every architecture, which I call the entry audit: when a predicted continuation can be read in two ways that call for different actions, what in the system selects the reading, and where did the selector come from? Drawing on the builders' own technical reports and on independent evaluations of their systems, it finds in every case a locatable import: curated data, probes trained on human labels, goal images, rewards derived from annotation, supplied camera frames, game engines, logged trajectories, or a language model. Part I sets out the general argument: the fibres of the observation map; the selection of data; the probe as the fixing of a gauge that identifiability theory leaves free; the difference between prediction and state; and the supplied goal. It proves what I call the hybrid paradox: a certified architecture never admits an action whose safety depends on a distinction its formal layer does not itself draw, so that a formal layer able to draw every such distinction would make the learned layer redundant for safety, while a learned layer that could be formally controlled would, for safety, be a formal layer under another name. It then shows, with a recent finite-testing result and a direct argument from underdetermination, why no campaign of tests in a simulator yields a bound, free of assumptions about the distribution the system will meet, on the cases in which simulator and world come apart, and that the tested cases are themselves a selection whose remainder no score measures. Part II audits each family through the achievements its builders report, among them an agent that learns from recorded play alone to mine iron in Minecraft and a driving simulator validated at three separate levels against real outcomes, conceding those achievements. Part III records what the variations change, proposes an ambiguity test that would discriminate between the claims, and states a falsifiable re-absorption prediction. The variations move the point at which meaning enters, so that it arrives in one family with the data and in another with the goal or the language model, and none of them removes that point.