Evaluation Sets That Lie to You

Placeholder article. Leakage rarely looks like leakage. It looks like a model that performs suspiciously well.

The common shapes

Random splits over data with temporal structure. Near-duplicates across the split. Normalisation statistics computed before splitting. Hyperparameters tuned against the set you later report on.

The fix is procedural

Split first, by the axis that matters — time, patient, site, document. Then do everything else. Touch the test set once.

Replace this placeholder article with your own writing.