A validation comparison uses five random seeds and reports a distribution. Why is this useful?

1. It makes every hyperparameter choice equally good
2. It eliminates the need for a held-out dataset
3. It proves all future deployments will match the mean
4. It exposes variability beyond a single favorable run