A neural network learns Pac-Man, a world model learns to dream the game, and a third agent learns to play entirely inside those dreams. PPO → RSSM world model → imagination training → online dream loop. Custom vectorized NumPy engine, PyTorch, built from scratch.