A fast leakage scan
- Check whether target-derived columns leak into features.
- Verify time-based splits when timestamps exist.
- Confirm feature engineering is fit only on training data.
If a metric feels suspiciously strong, leakage is often a better first hypothesis than breakthrough.