Bias, variance, and leakage in plain code — step 2 of 7

First training run on a new churn dataset: 99.8% accuracy on the held-out test set. What's the most likely explanation?