The policy engine.

What it is

A deterministic rule layer between your agent and the venue. Every action the model proposes is checked against your rules before it can execute. Same inputs, same decision — there is no model judgment inside the check.

The rules you set

• Loss limits — cap what an agent may lose. A proposal that would breach your limit is rejected, not debated.
• Asset allowlists — the markets an agent may trade. Anything off the list is rejected.
• Approval rules — the actions that require your explicit sign-off. Nothing on that list executes without your approval.
• Payload-bound approvals — an approval binds to the exact proposed action. A changed action is a new proposal, not an approved one.

What a rejection means

A rejected action does not execute, and the check is part of the turn record — you can read the why in the debrief.