[T15 · Bias-Variance] A team evaluates model performance across 5-fold cross-validation. The per-fold accuracies are: [0.82, 0.83, 0.91, 0.82, 0.81]. They report mean = 0.838. A statistician flags the fold-3 outlier. What might cause one fold to perform much better than the others, and what should the team investigate?