Researchers propose a novel approach to scaling world models by utilizing agentic game development as a verifiable trajectory data engine. The authors argue that current scaling strategies, which rely heavily on massive video datasets and fuzzy reward proxies like CLIP scores, are inefficient for spatial generation. To address this, they introduce Reinforcement Learning with Human-Engine Verification (RLHEV), a post-training paradigm designed to provide grounded reward signals. In this framework, a game engine acts as an executable world specification, allowing for the efficient verification of physical properties such as collision detection, physics accuracy, navigability, and bounded playability. Unlike unstructured video, game environments provide dense, verifiable signals that can serve as high-quality rewards for Reinforcement Learning (RL). By combining these automated engine-based signals with implicit human feedback during the development process, RLHEV creates a recursive data engine. This method not only generates long-horizon, real-world trajectory data but also provides the precise, non-fuzzy feedback necessary for the effective RL post-training of spatial world models, mirroring the success seen in code-based LLM training where compilers provide executable verification.