A team uses automated ML to compare candidate models. What must it still define and review?

1. Remove every independent test set
2. The task, suitable data splits, metrics, constraints, and final validation
3. Assume the highest training score proves production suitability
4. Give the search unrestricted access to production secrets