A model team shares results as screenshots without recording evaluation dataset versions. Which approach best fits?

1. Investigate leakage and spurious correlations with representative validation.
2. Record metrics with the exact evaluation data and procedure.
3. Evaluate a governed feature-management approach appropriate to batch and online access.
4. Unify and test transformation semantics across training and serving.