AI Layer
3. Policy Compiler

3. Policy Compiler

Risk rules in plain English, compiled to deterministic on-chain policy.

Write the rule in English

“Never trade NVDA within one hour of earnings.”
“Cap single-order size at 2% of pool depth when the NYSE is closed.”

The compiler translates these into deterministic, auditable on-chain policy for the ERC-4337 guard.

The hard boundary

The LLM writes the rule; it never executes the decision.

Enforcement stays deterministic. This separation is a hard architectural requirement — it is what makes it defensible to give the guard authority over real capital.

Model choice

This task runs on claude-opus-5, the most expensive tier in the system. It produces on-chain policy that executes money; wrong output means lost funds. See LLM tiering.