Abstract
This disclosure describes a way of packaging and installing AI software so that a model never runs apart from the rules and evidence that make it safe to run. A governed bundle contains, in one signed unit: the model's weights or a reference to them by hash; its birth certificate and lineage; its policy, such as the governor ladder, permitted and prohibited actions and data rules; its declarations of what it senses and what it may do; the conformance tests it was certified against and their signed results; and the certification class it holds. A certified runtime, such as a robot's controller, a personal AI computer or a data custody node, installs a bundle only if every part is present, every hash matches, the signatures are valid, the certification is in force, and the bundle's class is one the runtime is certified to host. At run time the runtime enforces the bundle's policy and declarations from its own safety plane, so that the model cannot run with a different policy than the one it was certified with. Updates arrive as complete new bundles, not as loose patches; a new bundle runs first in shadow beside the current one and takes over only after its own checks pass. Bundles are recorded in a public transparency log, so that anyone can verify which bundle a system is running. Any change to any part produces a new bundle that needs its own certification.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Ware™: A Governed Software Bundle for AI Systems, Binding a Model to Its Policy, Declarations and Test Evidence in One Signed Unit that Certified Runtimes Install Only Whole and Verified", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12236