Abstract
This disclosure describes a safety system for AI systems that joins the functions of starting, running, stepping down, stopping and restoring an AI system into one chain, published as an open reference design that any maker may build and have certified. Each AI system is divided into a mission plane, where the AI model and its applications run, and an independent safety plane, built on separate hardware with its own power, clock and channel. The safety plane, not the AI, owns every change of state in a declared lifecycle: off, initiating, running under governance, stepped down to a declared safe mode, held in a safe state, and reset and restored. Start-up passes only through a boot-start function that verifies the system and its parties; running proceeds only through a single governance point and a renewable power lease; step-down follows declared safe-mode levels; stop and restore follow an out-of-band reset. Each transition is signed and recorded, and the AI can neither block nor reverse a transition owned by the safety plane. Stop paths are controlled by the system's owner or operator and are opt-in: a physical key or button, a local channel, and expiry of the power lease. No third party, including the publisher of the design, can disable the system remotely, and the design contains no hidden access path. Systems declare a safety profile, basic, standard or advanced, according to risk, and each profile sets the required independence, diagnostics, response times and test coverage; profiles for machines that act in the physical world are aligned with established functional-safety practice. Every build carries a machine-readable safety case linking each safety claim to the tests and evidence that support it, updated with each release. The design, comprising hardware, firmware, interfaces and tests, is published openly; a build that matches the reference and passes its tests is certified by equivalence, and certified builds run continuous self-tests and periodic proof tests. A standard safety interface lets any AI model or runtime, including uncertified models hosted as contained guests, sit inside the chain. If the safety plane itself fails, the system goes to its safe state; if connection is lost, it continues at a reduced tier. Each safety system also acts as a node in an open, orchestrated safety network: it shares signed signals of faults and attacks, receives signed advisories and step-down instructions, and takes updates to its tests and profiles from a living standard, so that the safety of many AI systems rests on a continuous, layered and evolving network of mechanisms rather than on any single one, with no single point of control.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Safety System™: An Open, Certifiable Safety Chain for AI Systems, in which an Independent Safety Plane Owns Every Transition from Start to Running, Step-Down, Stop and Restore, with Owner-Controlled Stop Paths and No Hidden Access, Graded Safety Profiles, a Machine-Readable Safety Case, Builds Certified by Equivalence to an Open Reference, and Each System Acting as a Node in an Open, Orchestrated Safety Network", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12626