A model registry must distinguish approved production versions from experimental artifacts. Which approach best fits?

1. Implement a versioned ML pipeline.
2. Use valid step caching keyed to the relevant inputs and implementation.
3. Add an explicit evaluation gate before deployment.
4. Track model versions and promotion state with accountable approval.