1. Context retrieval is not execution.
Reading the mission, repository, memory, or runtime can resolve what is true. It does not mean anything changed. George keeps that boundary explicit.
The sequence above is the short version. Here is the machinery underneath it.
Human intent enters through ChatGPT reasoning and continuity. George provides governed ingress, semantic mission/state, authority and contract admission, executor selection, evidence normalization, verification, and the next justified transition.
Reading the mission, repository, memory, or runtime can resolve what is true. It does not mean anything changed. George keeps that boundary explicit.
A model can reason about a change without having authority to make it. Planning authority and mutation authority are separate facts.
Starting a worker, command, deployment, or external action says only that work was launched. George retains the operation identity until the intended effect is actually known.
A long operation can still be the right operation. Timeouts and reconnects do not automatically justify launching it again. George reads back the same operation when the work is still in flight.
A connector failure is not automatically a George failure. An executor failure is not an evidence failure. A stale test is not proof that current runtime behavior is wrong. Classification matters because the next repair depends on what actually broke.
MCP is governed ingress, not the identity of George. Executors, builders, connectors, processes, and capabilities can change while semantic mission identity, authority, evidence, and open obligations remain above them.
A worker saying “done” does not close the mission. Stronger runtime, repository, test, video, or primary evidence can reopen the question and invalidate a previous conclusion.
The work remains itself even when the machinery underneath it changes.