Abstract
This disclosure describes a way of grading AI systems against a standard so that the grade always reflects what the system can prove now, not what it once passed. The standard defines conformance levels, each a cumulative set of safeguards: for example, identity and records at the first level; a runtime governor and declared policy at the second; an independent reset, independent sensors and a power lease at the third; and live certification, incident reporting and continuity at the highest. Each safeguard is defined together with automated tests that prove it, such as injecting a forbidden action and checking that it is refused, cutting the safety channel and checking that the system stops, or presenting an expired mandate and checking that it is rejected. The tests run continuously in the field, at declared intervals and at random, on the system itself or on a test harness beside it, and their signed results flow to the certification body. A system's level is carried in its passport. When a test for a safeguard fails or is not run on time, the level drops automatically to the highest level whose safeguards are all still proven, and hosts that require a higher level stop admitting the system at their next check. The level is restored when the failed test passes again. Domains such as healthcare, transport or homes add their own profiles on top of the levels, and certification bodies that run the same tests recognize one another's grades.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Standard™: Graded Conformance Levels for AI Systems, Each Defined by Safeguards Proven by Automated Tests that Run Continuously, with the Level Carried in the System's Passport and Lowered Automatically When a Test Fails", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12237