Abstract
This disclosure describes how AI systems that are not certified can still be used safely, by running them as guests inside a certified hosting environment. The environment, such as a runtime on a device, a server or a site, is itself certified, and declares a compatible status in its own passport. An uncertified model placed in it runs inside an isolated compartment with no direct access to the host's network, sensors, actuators, data or keys. Every input to the guest and every output from it passes through the host's governance point, which applies the host's own policy, inference and output contracts, and governor levels, so that the guest can propose but only the host can act. The host limits the guest's reach by declared scopes, such as read-only access to stated data, no actions on anyone's behalf, no communication with other guests or devices, and budgets for compute and traffic. All output that leaves the host carries a machine-readable label naming the guest as uncertified, with its identifier if it has one. The host's safety plane can pause, reset or remove the guest at any time, and the guest is included in fleet-wide resets through its host. The host records what each guest did, and evidence from that record can support a later certification of the model. Certified systems and spaces that admit only certified parties can admit a compatible host, and through it the guest, within the guest's declared scopes.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Compatible™: A Certified Hosting Environment in which Uncertified AI Models Run as Contained Guests, Wrapped in the Host's Governance, Limited in Reach, Labelled in Their Output, and Reset through the Host", Technical Disclosure Commons, (November 02, 2026)
https://www.tdcommons.org/dpubs_series/12614