Abstract

This disclosure describes an open way to add applications to certified AI systems, such as home assistants, robots and agents, without a single gatekeeper and without weakening the system's safety. Each app is a governed software bundle with a manifest that declares, in machine-readable form and in plain language, the actions it can take, the data and sensors it uses, the limits it needs and its purposes. When an app is installed, the host system checks its signature and certification status through the passport handshake, and the governor admits it only for capabilities that the system's own certification and the user's settings allow. Permissions are granted as expiring, scoped grants that the user can see and withdraw, not as permanent rights, and an app can never raise its own permissions. While it runs, the app is confined to its own compartment and every action goes through the governor, which compares its behaviour with its manifest; behaviour outside the manifest raises the governor level, suspends the app and produces an incident report. Apps can be listed in any number of catalogues run by anyone; trust comes from the app's signature, certification and live status in the registry, not from the catalogue owner, and each listing shows that live status, including recalls. Recall is built into the host and the registry, not the app: every update is admitted afresh, a modified app is refused, and an app found rogue after launch is recalled from every installed copy, whatever catalogue it came from. Because apps call a common set of governed calls, the same app runs on certified systems from different makers. Uninstalling an app withdraws its grants and erases its data with proof.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS