A training job runs correctly locally but fails because a library version differs in OCI. What should be recorded and deployed?

1. Only the source notebook's filename
2. A reproducible environment containing the tested dependency versions
3. Only the model's prediction threshold
4. The latest version of every library on each run