The visible output may look like video, but the product requirement resembles a game loop: input must affect what appears next quickly and consistently.
Interaction is the boundary
A video model can encode physics without exposing live control. We classify a system as Class C only when user or agent actions affect future generated observations during the running session.
What to measure
Action-to-effect latency, accepted input frequency, output FPS, control vocabulary, session horizon, object state, spatial memory, and revisit consistency determine whether a demo can become a product.
Where it may be used
Generative games, robotics simulation, embodied-agent training, interactive stories, virtual production, and persistent live worlds are early application categories.