RP-Worlds

World growth — new operable entities appearing during simulation — has so far been treated as unconstrained generation rather than as a validated state transition. LLM-agent worlds are structurally fixed: the entity set is frozen at construction, so the state may evolve but the world cannot grow; free-form LLM worlds do grow, but, lacking a validated canonical state, they drift. We introduce the RP-world: play-centered agentic world simulation in which interaction expands the canonical world itself. A world-agent — the single writer of a validated tree of typed entities, formalized as a Parameterized-Action POMDP — plays opposite M soul-agents that own only their own minds. As play pushes beyond the map, the world-agent proposes new places, inhabitants and relations and admits them only after commit-time validation, so the canonical state provably cannot accumulate contradictions. Across six story-complete worlds and three model families: validation eliminates state contradictions where the otherwise identical architecture without it drifts universally; effective state dimensionality (ESD), our measure of growth, exceeds the entity ceiling of fixed-extraction frameworks while contradictions stay at zero; and a transcript-only LLM judge still prefers the drifting baseline — fluent prose conceals state drift. Drift is architectural, not a model limitation.