LLM Behavior Diff Model Update Detector

You swap a model. The new one scores better on your benchmarks. You deploy it. Two days later, a user reports that something that used to work reliably now behaves differently. The benchmark never cau