Abstract
This disclosure describes a coordination layer for fleets of AI systems and the equipment they operate, designed so that central coordination improves safety without becoming a single point of failure or a single point of attack. The layer has a front end and a back end. The front end is the owner's console: for each system or fleet, it shows the current start statement (what the system may do today, under which safety profile, and how to stop it, with the stop shown first), live status, declared limits, any step-downs or safe modes in force, quarantine notices, and the signed record, so that the owner's authority over AI is visible in one place. The back end is a control plane that connects the stages of the safety cycle (start, check, run, govern, stop and restore) through open protocols: it distributes policies and limit sets, delivers signed advisories and quarantine notices to every connected system, and broadcasts safe-mode instructions. Its powers are asymmetric by design. It may tighten limits, enter safe modes and isolate agents at once; it may never loosen limits, lift a stop, restart a stopped system or override a system's own safety controls. Loosening requires delayed approval by two accountable parties, applied at the system itself. The layer is never in the stop path: each system's owner stop works through its own path whether or not the coordination layer is reachable. Every message from the layer is signed, and each system checks it locally against its own limit set and refuses anything unsigned, outside its limits or inconsistent with its profile, even when it appears to come from the layer. If the layer fails, is unreachable or is suspected of compromise, every system holds its last safe state and keeps running, and the layer's keys can be revoked by the owners without its cooperation. The layer has no hidden access, and every action it takes is written to the same signed registry that governs the systems and serves a public lookup within privacy limits.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Orchestra™: A Coordination Layer for AI Systems that Joins the Owner's Console to a Control Plane across Every Safety Stage, Can Tighten Limits at Once but Never Loosen or Override a System's Own Safety, Is Never in the Stop Path, Sends Only Signed Messages that Each System Checks Locally, Leaves Every System Holding Its Last Safe State If It Fails or Is Compromised, and Records Every Action in One Registry", Technical Disclosure Commons, (November 09, 2026)
https://www.tdcommons.org/dpubs_series/12843