2 quick questions

Hi! I hade 2 quick questions: First, about the overfitting test: suppose a predictor's train() is effectively a no-op and predict() computes a fixed transformation of the current features, so retraining on permuted labels gives the same predictions. Is that kind of fixed, label-free construction eligible, or does the test require the model to depend on the training labels? Second, about the transaction cost penalty mentioned in the profit sharing thread: will it apply to this competition's leaderboard and final scoring, or only to the onboarding selection, and from when?