Each example’s weight varies with its predicted probability, decreasing as the model becomes more confident.