What the Split Does and Does Not Say - Train/Test Split, Exactly

Close the boundary around one exact split and one exact held-out score.

highlighted = computed this step

The no-leakage boundary

The train statistic uses train rows only. The test rows are held out and scored once.

\text{train first; score test once}

What this does not promise

This split does not prove generalization, future performance, unbiased measurement, safety, or the best split.

\text{split arithmetic only}

The honest close

This book pins one split and one constant baseline. It is NOT training, NOT learning, NOT generalization, NOT future performance, NOT unbiased, NOT safe, NOT best split, and NOT an accuracy guarantee.

\text{held-out rows stay held out}