Abstract
This disclosure describes how a standard for AI systems can keep pace with the systems it governs. The standard is written as numbered clauses, each a requirement in plain language bound to one or more executable tests and to the hazards it addresses. Clauses and tests are versioned together, and every system's certification states the version of each clause it meets. Changes follow two tracks. On the ordinary track, proposed clauses are published with their tests, tried on volunteer systems, opened for public comment and adopted on a published schedule, with a declared transition period before they become required. On the emergency track, when incident reports or field tests reveal a new hazard, an emergency clause and its test can be issued within days: the clause is signed by a quorum of independent reviewers, published with the evidence that prompted it, and delivered to certified systems in the field as an update to their continuously run tests. Systems that cannot pass an emergency clause by its deadline have their conformance level lowered automatically. Each emergency clause expires unless it is confirmed through the ordinary track within a set period. Systems certified against an older version remain valid until their transition period ends, and every system's passport shows which clauses it meets, so hosts can require specific clauses for specific uses. Retired clauses and their tests remain published, so the history of the standard is permanent.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Standard™: A Living Standard for AI Systems, Written as Versioned Clauses Each Bound to Executable Tests, with Emergency Clauses Issued from Incidents and Pushed to Systems in the Field within Days, and Declared Transition Periods", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12278