A pipeline reruns unchanged deterministic preprocessing on the same versioned input and configuration. Which approach best fits?

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