Two runs have different validation results. Which evidence is needed to interpret the difference?

1. Only the notebook's wallpaper
2. Only the model's display-name length
3. Training code, data versions, environment, parameters, and evaluation procedure
4. Only the final scalar score