"World model" has become one of the most stretched terms in AI. It is used for video generators that produce convincing clips, for 3D scene reconstructions, for robot policies and for interactive game-like simulators. A new preprint from OpenEnvision, a research collective of 16 named authors, argues that the field cannot measure progress until it agrees on what the term means and runs these systems on shared plumbing. Its answer is WorldFoundry, open-source infrastructure for running and scoring them.
The paper starts with a definition. A generative world model, in its terms, keeps a persistent internal state of an environment, updates that state when someone acts on it, and produces coherent futures that reflect those actions, including ones that did not happen. A video model that makes pretty footage but forgets the room's layout when the camera turns does not qualify. The authors argue that visual plausibility is neither necessary nor sufficient evidence of a real world model, a gap that physics-focused benchmarks keep exposing.
The software is the practical part. WorldFoundry brings model and benchmark discovery, environment setup, execution, artifact storage and visualisation under one roof, with adapters for video, 3D and 4D, embodied action and interactive systems so that each keeps its native inputs and evaluation rules. Its main design choice is to separate whether a model actually ran from how well it scored: manifests declare what each entry needs, outputs are stored behind fixed contracts, and scorecards record execution evidence apart from performance claims. The current catalogue lists 270 model entries, 74 benchmarks and 62 metric entries, and the team says every model runner has been checked end to end against a documented reference workflow.
This matters because world-model results are notoriously hard to compare. Labs each use their own setup, and a demo video often stands in for evidence. A shared harness of this kind could do for world models what common evaluation tools did for language models, making it easier to check a claim from a system such as Nvidia's Cosmos family against rivals on equal terms.
The caveats are about adoption and depth. A catalogue entry that runs is not the same as an evaluation the original authors would endorse, and the paper's definition, while useful, is the authors' own and will not suit everyone. Infrastructure only pays off if other labs use it, and the preprint cannot show that yet. It has not been peer reviewed.