A team compares results from two model versions. What makes the comparison most interpretable?

1. Different unseen datasets selected separately for each model
2. Only the model names sorted alphabetically
3. Selecting each model's best training batch score
4. The same held-out dataset and explicit metric definitions