Metric Is Not Generalization - Evaluation, Honestly

The capstone separates exact measurement from broader claims. Metrics can be computed honestly on held-out data without promising that future rows behave the same way.

highlighted = computed this step

Metrics measure shown data

A metric is a measurement on the rows it was computed from. The exact fractions in this book are useful because the numerator and denominator stay visible. They are still measurements of this held-out set, not promises about future rows.

\text{metric}=\text{measurement on shown rows}

The claim boundary

The rates can be exact without becoming a claim about future rows. The same fraction can look strong on one split and weak on another. Honest evaluation keeps the data source, threshold choice, and denominator attached to the number.

\text{exact metric}\ne\text{future guarantee}

Summary

This book pins the mechanics of evaluation: held-out rows, chosen thresholds, exact counts, and exact rational rates. It does not claim training, does not claim the right threshold, and does not claim the model works on new data.

\text{measurement, not a promise}