Can a World Model Live Partly Outside the Weights? Persistent Developmental State for Continual Scientific Reasoning

Continual adaptation may have more than one substrate. We present an implemented local research architecture in which a language-model renderer $W$ is conditioned by an explicit persistent developmental research state $S_t$: outputs are produced as $y_t \sim P_W(y \mid x_t, S_t)$, while the state itself updates selectively as $S_{t+1} = U(S_t, E_t, A_t)$ — without changing the model weights. Across successive implementations of this architectural lineage, we observed persistent consequence traces, developmental continuity across three research modes through a shared persistent state, lineage-preserving continuation across heterogeneous models, operating systems, and machines — and two documented episodes of governed, evidence-based revision of the system's own research strategy. We distinguish developmental state from retrieval memory: it is selective, revisable, lineage-aware, consequence-coupled, and conditions later research rather than merely storing it. We then propose three falsifiable tests — state intervention, heterogeneous inheritance, and state prediction. The implemented system demonstrates the substrate; prediction of future state transitions remains the missing test for a stronger continual-world-model interpretation.