Split one accuracy score into a confusion matrix's four cells, and weigh a false alarm against a missed case to see why a 99%-accurate model can be useless on rare events.

A screening model looks at each patient and, every time, returns `healthy`. It is right for 999 of every 1,000 people, a score most teams would ship, and it has never once caught the patient who is actually sick.

Compare two models fairly by rotating the held-out slice instead of trusting one lucky split, keep the simpler model when scores tie, and seal the test set so checking it never inflates the verdict.

You ship Model A over Model B because A scored higher on your test pile. A teammate reruns the exact comparison with `random_state=3` instead of `0`, B now wins, same code and same emails.

Trace a model's unfair tilt back to the skew in its training data, and judge when a confident prediction has not earned your trust.

A résumé scoring model trained on a decade of mostly-male hires reads a new application and returns `reject`. No one wrote that rule; the model only ever learned the one pattern the old hiring data already carried.