Evaluation starts by naming which rows are held out. This lesson keeps the split as exact integer bookkeeping and names leakage as the failure mode where the boundary is broken.

highlighted = computed this step

Hold out evaluation rows

Evaluation needs a held-out set: rows not used by the fitting procedure. In this toy split, there are 10 examples total, with 6 kept for training mechanics and 4 held out for evaluation. The split is an integer bookkeeping choice before any metric is read.

N=10,train=6,test=4N=10,\quad \text{train}=6,\quad \text{test}=4

Name the failure mode

Leakage means test information slips into the fitting procedure or threshold choice. Then the evaluation set is no longer held out, so the reported metric is measuring a contaminated process. Naming leakage keeps the measurement boundary visible.

leakage=test information enters the fit\text{leakage}=\text{test information enters the fit}

Summary

The held-out split is exact integer accounting, not a performance promise. It says which rows are measured after fitting mechanics stop. It does not claim that future rows behave the same way.

held-out counts first; claims later\text{held-out counts first; claims later}