Train models sequentially so later models focus more on previous errors.

Replace all features with missing values.

Use no evaluation metrics.

Train only one model and never update it.