Skip to main content
Predicates are evaluated outside the LLM; the model cannot bypass them. Each policy predicate is a boolean rule enforced before any transaction is signed.

​
System diagram

Settlement happens on Mantle Sepolia; reasoning traces anchor to 0G Storage.

​
Predicate set

​
Composition

Additional runtime checks in GuardrailService.check_plan: observation quality, balance sufficiency, gas spike guard, protocol whitelist.

​
Failure mode

Any predicate false → plan rejected → guardrail_evaluated event with violations[] → surfaced in replay Policy validation node. No on-chain execution proceeds.

​
Design principle

Most agent stacks treat policy as an LLM judgment call. AMEO separates cognition (LLM or rules planner) from authorization (deterministic predicates). A fluent rationale cannot override a failed drawdown or whitelist check. To add a custom predicate, see Write a policy.