A team tunes hyperparameters repeatedly using the test set. What is the main concern?

1. The training set automatically disappears
2. Every model becomes perfectly calibrated
3. The model's parameter count must become zero
4. Evaluation begins to overfit the test set