In this competition, predicting the correct class is not enough.

Because submissions are evaluated using **Multi-class Log Loss**, models are rewarded for:
- Assigning higher probability to the correct action
- Avoiding overconfident wrong predictions

### Best Practices
- Use validation sets to evaluate calibration
- Avoid hard 0 or 1 probabilities
- Consider techniques such as:
- Temperature scaling
- Platt scaling
- Regularization and early stopping

In real decision systems, **confidence quality matters as much as accuracy**.