AI Control
A research programme that treats a capable model as a potential adversary rather than a tool, and asks what a deployer can guarantee about a system it cannot inspect. The unit of analysis is the protocol — the arrangement of trusted and untrusted models, monitors, audits and access rules — not the model's weights. Its defining move is to assume the model may be actively trying to defeat the protocol and then measure how much damage it can do anyway.
Why It Matters Here
This is the field's name for what the corpus argued from fiction. The Operator's central claim — that reachability, not strength, is the design variable of a binding — is the control agenda's premise, reached independently and stated in different vocabulary. See CNF-001.
The difference is that control research measures the assumption. It reports which protocols hold at which capability ratios, using red-team games where a model is instructed to attack the protocol and its success rate is counted. The corpus asserts the ordering; control research prices it.
What It Adds to the Corpus
- A capability differential as an explicit design variable. The corpus's layers have no column for the relationship between binder and bound; control research makes that relationship the whole question.
- Trusted editing and untrusted monitoring as named protocol elements, which are more specific than the corpus's "oversight" layer.
- An admission that the protocol can be defeated and a method for finding out how — which the corpus's design rules do not have.