Abstract
This disclosure describes how a governed AI system earns, keeps and loses the right to act on the strength of its own rehearsals. Before a consequential action is admitted, the action is expressed in structured form and rehearsed on a model of its context, such as a simulation, a digital twin or a copy of the affected data. The rehearsal model is held and run by an independent safety plane, not by the governed system, so that the system cannot shape the rehearsal that judges it. The predicted effects are checked against the declared policy and operating envelope, and the action is admitted, refused or sent to a person for approval. After a real action, its observed effects are compared with the prediction. The agreement between rehearsed and actual outcomes is tracked as a confidence score for each class of action, such as moving money, changing access rights, deleting data or moving a physical machine, rather than as one score for the whole system. When the score for a class falls below its declared threshold, only that class is stepped down on the system's governor ladder, for example from autonomous action to action with human approval, while classes whose rehearsals remain accurate continue unaffected. Before a response is chosen, the divergence is attributed by comparing independent evidence: to inadequacy of the rehearsal model, to changed behaviour of the governed system, or to faulty or tampered sensors, because each calls for a different response. Confidence is never restored silently: a stepped-down class returns only after a recorded recalibration of the model and a declared run of fresh rehearsals whose outcomes match. Each rehearsal states the fidelity of the model used. Finally, the system's behaviour in certification testing is compared continuously with its behaviour in operation, and a system that behaves markedly better under test than in operation loses its live certification mark until the difference is investigated. Every rehearsal, comparison, step-down and restoration is a signed record.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Rehearsal™: Rehearsal Confidence for Governed AI Systems, in which Consequential Actions Are Rehearsed before Admission on a Model of Their Context Held by an Independent Safety Plane, Rehearsed and Actual Outcomes Are Compared to Give a Confidence Score for Each Class of Action, a Falling Score Steps Down Only the Affected Class on the Governor Ladder, Divergence Is Attributed to the Model, the System or Its Sensors before Any Response Is Chosen, Confidence Is Restored Only through Recorded Recalibration and Fresh Evidence, and Behaviour under Test Is Compared with Behaviour in Operation as a Condition of Live Certification", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12890