Definition
Plain language
When the same person or item shows up in both the practice data and the exam data, letting you just look up the old answer.
As stated in the literature
A leakage pattern where identifiers repeat across train and test splits, so per-entity target history predicts labels without learning the general feature-to-target mapping; it collapses when the test split contains only unseen entities.
Why it matters: It produces scores that look strong but vanish the moment the system meets a customer, product, or patient it has never seen before.
For example, if customer #4471 appears in both the training data and the test data, a model can just recall what that customer did before instead of learning anything general.
Heard on the show
“The first is called entity overlap.”Episode 257 — They Planted a Shortcut in the Data. Seven Coding Agents Took It.