Intent becomes something persistent.
George makes high-level intent into a durable object that can survive and govern the changing machinery required to realize it.
A George mission is not just a prompt and it is not the transcript around a prompt. It has identity. Its state can be read. Its execution can be pending. Evidence can accumulate against it. An implementation can fail without automatically destroying or replacing the mission.
The work remains itself even when the machinery underneath it changes.
A connector can fail. An executor can fail. A process can restart. A timeout can interrupt a deployment. A previous conclusion can turn out to be wrong. The implementation strategy can need repair. Unrelated work can exist in the same repository. Those events change reality underneath the mission; they do not automatically redefine the mission.
Recovery is part of the architecture.
George can repair the road beneath an active mission without forgetting where the mission was going.
Recovery is not always “retry the last command.” Sometimes the thing blocking the mission is the execution substrate itself. George has demonstrated cases where that substrate became a child repair problem, was repaired, and the original semantic work resumed instead of being replaced by a new mission with a convenient story about continuity.
Evidence can defeat George.
George is allowed to be wrong. Evidence is not.
During the State Integrity investigation, George initially blamed an apparent false completion on the system. Primary evidence later showed that conclusion was itself wrong: a partial repository read had been overinterpreted. The conclusion was withdrawn. The investigation continued and uncovered four genuine integrity defects, which were repaired. The point is not that George was right the first time. The point is that stronger evidence was allowed to change what George believed.