A deployed churn model computes a customer's average weekly usage using one implementation during training and a separately written implementation in the prediction API. The two implementations handle missing weeks differently. What deployment risk does this create?