A model release can contain code, weights, documentation and datasets under different grants. Treat them as separate artifacts. Availability to download does not establish approval under an open-source definition or permission for every use.
Use a version-specific license. Llama 3.1 includes attribution and acceptable-use conditions and additional commercial terms tied to licensee scale at release. CreativeML OpenRAIL-M has its own use restrictions. A research-only weight grant does not become commercial merely because an inference wrapper is MIT.
Training-data provenance and the legal status of a generated output cannot be inferred from the weights' license alone. A coding agent's assertion that it generated a snippet also does not prove there is no upstream copyrighted expression.
The model exercise supplies the use review and scale facts so the remaining attribution obligation is checkable. The provenance drill uses a known upstream source and a supplied grant; it does not attempt probabilistic plagiarism detection.