In this work we evaluate a structured multi‑agent world model from the outside, covering a full ten‑hero MOBA (206 units, every hero acting each tick, games up to 6,000 ticks). The policy is trained solely inside the model using 1,400‑tick free‑running imagined episodes, then measured against the shipped opponent in the real game. The real game supplies no gradient; it only provides the policy’s own matches as training data for the world model and an online evaluation that anchors the policy. Running as a continuous asynchronous Dyna loop, the policy wins 70.2% of real games as radiant (421/600, 95% CI 66.4‑73.7) on seeds never used for decision‑making, while it loses every game as dire and the shipped opponent also loses when playing itself. Four key findings explain the outcome: (1) model exploitation is invisible inside the dream—unanchored runs collapse after a few updates with no in‑dream metric detecting it; (2) a model accurate on its training corpus is severely wrong on the policy’s own games, and Dyna repairs this, adding roughly +9.2 % to the real win rate; (3) the policy inherits the model’s fidelity mechanic by mechanic—the model captures macro‑game dynamics but not crowd control, cast timing, or lethality, so the policy wins by map‑wide pressure rather than coordinated fighting. We release the world model, the dream‑PPO harness, a world‑model debugger, the evaluation protocol, and all policies and logs.
Review: This study convincingly shows that external evaluation combined with a continuous Dyna loop can turn a purely imagined‑training policy into a strong competitor in a complex multi‑hero arena, and it provides a fully reproducible benchmark and tooling for future research.