A model performs poorly after timestamps are parsed in a different timezone during inference. What should be standardized?

1. Only the final probability rounding
2. A separate arbitrary timezone for every request
3. Only the model catalog display name
4. The shared training and inference timestamp transformation contract