AI models can ace every benchmark and still cause harm. Here is a practical, layered way to measure safety, robustness and governance, not just accuracy.