Systematically testing a trained model to find bugs, shortcuts, or unintended behaviors before deployment.