A team needs repeatable preprocessing, training, evaluation, and registration steps with tracked dependencies. Which approach best fits?

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