AI that weighs risk instead of blocking

Most chains treat a transaction as binary: let it through, or reject it. A single threshold has to serve both the obvious attack and the slightly unusual payment, and it serves neither well.

The Sentinel Policy Layer replaces that with a graduated response to the risk of each transfer:

• low — goes through immediately
• elevated — a short delay with a window to cancel
• high — requires confirmation with the owner's master key

The part that matters for agents

An agent can initiate a transfer. A risky one is confirmed only by the owner's master key — an agent's session key physically cannot approve its own operation. That is not a policy an application enforces; the ban lives in the authorisation path itself.

It means you can delegate real spending authority to software without delegating the ability to escalate.

Deterministic, on every node

All of it runs at the consensus layer, so the same transaction is scored identically everywhere. Nothing here depends on trusting one node's judgement.

AI-native security is not a service running on top of the chain. It is part of the protocol.