In Human Compatible, you laid the groundwork for "provably beneficial" AI, emphasizing the need for a definitive "off-switch." We believe the underlying truth of this "off-switch" must be realized physically, completely independent of the statistical model it governs.
Our vision at the OHM NI Stack is to solve AI security in a foundational, best-practice way where the entire academic community can contribute. We have filed USPTO Applications #63/994,444, #63/997,472, and #64/1,997,472 detailing the TLA (Monotonic Risk Ratchet). It serves as a mathematically enforced off-switch that physically severs execution pathways when deterministic constraints are breached.
Before we present this to the corporate world at the Planetary AI Safety Summit in Vienna (May 2026), we want CHAI to contribute to its validation. We invite you to brutally stress-test our implementation. We do not want this to be a closed ecosystem; we want it to be the verifiable best practice that every researcher can build upon. Will you help us evaluate this underlying truth?