leakage
When the test set shares information with the training set, so the test score is too good.
Leakage occurs when test examples are not independent of training examples (overlapping windows, the same session or person on both sides), inflating the measured performance.
Examples
Cutting one recording into windows with 50 % overlap and splitting at random puts half of some test windows' samples in the training set.
Training on Monday's sessions and testing on Tuesday's, or on a different person, keeps one sitting's grip and angle from leaking into the test.
See also