What should an experiment tracker record to make an observed improvement reproducible?

1. Only a screenshot of the final loss value
2. Only total GPU purchase cost
3. Data version, configuration, code, model and metrics
4. Only the researcher's preferred model name