for each level of media richness, there is a deterministic scaffold we lean on until the generative layer is fully capable on its own. for text, it was the python interpreter that stopped llms from hallucinating math. now that video conditioning is a thing, it is 3d rendering engines that not only enable control but also address the long-standing spatiotemporal consistency problem in world modelsthe question is when this scaffold should go away, if ever (python is still here even as llms got sma