Robustness and Distributional Shift

Why models fail when deployment data differs from training.

What is 'distributional shift' and why does it matter for safety?