Policy engine and audit log in development

The most important line of our architecture is where the model ends and the rules begin. Language models negotiate, draft and decide what to ask for. Whether money moves is decided by a deterministic policy engine that does not care how persuasive the prompt was. A manipulated agent can request anything, it still cannot spend past its mandate.

The second pillar is the audit log. Every intent, every policy decision, every execution lands in an append only log with hash chaining, so nothing can be quietly rewritten. Accountants get exports, auditors get evidence, and when the EU AI Act asks how an autonomous system is supervised, the answer is a file, not a promise.

We are building both pieces rail agnostic. The engine and the log do not know which bank sits underneath, which means a second banking rail can dock later without touching the control layer. That is deliberate: the moat is the control layer, and it should outlive any single integration.