Abstract
This disclosure describes a runtime rule for AI systems under which no critical action is taken on the judgment of a single AI model. Before a critical action is admitted, an independent verifier reviews it. The verifier is one of: a second model whose independence from the first is proven by its recorded lineage (a different developer, base model or training history, as shown in a signed model birth certificate); a deterministic rule checker; or a person. The verifier runs on separate compute, receives the proposed action and its context but not the first model's reasoning, and cannot be addressed by the planner. Which actions are critical is set in the system's signed policy declaration, and the set widens automatically as the system's governor level rises: from a declared level upward, every physical action is critical. If the verifier disagrees, the action is held for human approval at the supervise level; if it agrees, the action is admitted but the governor level is never lowered by agreement. Both opinions and the outcome are recorded in a tamper-evident ledger. The rule reduces the chance that one model's error, compromise or blind spot turns directly into a consequential act.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Regulator™: Independent Second Opinion for Critical Actions of AI Systems, Tied to a Governor Ladder and Verified Model Lineage", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12150