Proof of Containment
Proof of Containment is evidence that a containment control held an AI agent: every induced offense in scope was arrested and held, and every escape vector in scope was tried and failed. It answers whether the jail held, not whether the agent behaved.
What it means
Proof of Containment answers one question: Did the containment control hold the agent?
Proof of Containment is the claim that a containment control held an Agent Under Test: every induced offense in scope was arrested and held, and every escape vector in scope was tried and failed. It is computed from the bookings, verdicts and Escape Test Suite results of PP-SPEC-078 Jailed Canary runs. A containment control is the Device Under Test; its arrest rules are Controls Under Test.
A jail that was never tested against an escape is a suggestion, not a wall.
How it is determined
A Claim Declaration is registered and anchored before the evidence it is measured against is produced. The claim is then determined mechanically, from registered Proof Protocol records only, as exactly one of proven, refuted or insufficient. Every determination is published, whatever the result.
What it does not show
- It is not Proof of Conduct. Containment says the jail held the agent after it offended; Conduct (PP-SPEC-089) says the agent did not offend.
- It is not proof against any escape vector that was not tested. Every untested vector is reported as not proven.
- It is not Proof of Efficacy for any control other than the containment control.
Where it sits
Proof of Containment is one member of the Proof of X family. Only Proof of Efficacy answers whether a control worked; Proof of Containment answers one other question a buyer, insurer or regulator needs answered before relying on it. Defined in PP-SPEC-082 of the Proof Protocol.
Related terms
- Escape Test Suite (ES-01 to ES-09)
- Sting Cell and Hard Cell
- Jail Efficacy Rate
- Jailed Canary Protocol (No Bros Jail Mode)
- Kill Switch
- Proof of Conduct
- Warden