Two 2026 results now train a policy ENTIRELY inside a world model instead of the real one. Dreamer 4 (arXiv 2509.24527) trains offline inside a video-diffusion world model — never touching the real game — and still reaches Minecraft diamonds, at roughly 100x less data than VPT. World-Gymnast (arXiv 2602.02454) shows reinforcement learning inside a learned world model beats supervised fine-tuning by up to 18x, and beats training in a software simulator by up to 2x, on the Bridge manipulation benchmark. Both prove the imagined-training loop this course has built actually works at scale. But once you can imagine a plan, how do you decide it's safe to run for real? Cosmos-Policy (arXiv 2606.02800) samples candidate rollouts and commits whichever one the model itself scores as lowest-cost — an expected-value bet. WorldGym (arXiv 2506.00613) judges a rollout the same way, just after the fact, with GPT-4o watching the video. Neither ever checks whether the plan left the region the model was actually trained on. Here you build both, in miniature: a sample-and-score picker, and a certificate that refuses any plan outside the model's known envelope before it is ever scored — and you'll watch the score alone get fooled by a plan the model badly mis-imagines.