A patent pending on governing AI
A deterministic layer between every model and its user: an answer cites its source or it is refused.
The problem
Two failures pushed me here. The first was hallucination: model output you could not trust, because nothing checked it against a source. The second was a jailbreak: prompt-level guardrails turned out to be a suggestion, not a control.
Both failures have the same shape. The model is the last thing to touch the answer before the user sees it. Whatever you ask it nicely to do, it can be talked out of doing.
What I filed
On 23 August 2026 I filed a provisional patent application with IP Australia: AU 2026907221, A deterministic governance layer for large language model calls.
The idea is a control layer that sits between every model and its user. Every output is screened and grounded before release. An answer must cite its source or it is refused. The rules are enforced by the layer, in code, so they do not depend on the model's goodwill.
How it was hardened
The layer was not designed on paper. It was hardened over more than two years and more than 500,000 automated strategy development and review cycles, with 4,000 optimisation trials per cycle, inside my trading research. Every failure it let through became a rule. It now runs in production.
Why patent it
AI governance is mostly policy documents and prompts. Both are statements of intent. A deterministic layer is a mechanism: it holds whether or not the model cooperates. That difference is what I filed on, and it is the difference between hoping a system behaves and being able to show that it must.