Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models

Argues that scaling world models on crawled video is compute-inefficient and proposes a recursive data engine where game development acts as a reward environment: a game engine encodes an executable world specification that can check collision, physics, navigability, and bounded playability, with developers’ accept/reject decisions supplying a global verification signal. Introduces Reinforcement Learning with Human-Engine Verification (RLHEV), combining dense engine signals with implicit human feedback gathered during development.