Every agent run separates into two things. A decision layer that chooses the route, the model, the provider, the reroute. And a receipt layer that records each of those choices as a signed artifact. Deliver the receipts at the end and you get an audit. Stream them during the call to an independent dashboard and outside review can correct, reroute, or pause the trip while it is still moving.
Objective, plan, model, provider, reroute. The control plane over inference. When one provider is cheaper, faster, or a better fit for a task class, the agent can reroute — the same way a driver changes roads without changing the destination.
Each decision is canonicalized, hashed, and signed with ML-DSA-65. Receipts chain by prior-hash into one record of the whole trip. Stream them live and an independent observer holds the same evidence the agent does, at the same moment.
One button runs a short trip of seven route decisions. Each event streams in, then mints a receipt: a real ML-DSA-65 (NIST FIPS 204) signature over the canonical decision body, chained to the one before it, and verified in your browser against a keypair minted in memory for this run. Nothing is stored. Nothing is sent to a server.
———@noble/post-quantum ML-DSA-65 build this site uses on
/typed-signer/. Keypair, signatures, and verification all run in your browser.
The control-plane idea is the point: an agent that reroutes inference between providers when one is cheaper, faster, or a better fit — and signs a receipt for the decision — is the control plane over inference. This page renders that route and proves the decisions. It does not claim the underlying provider calls ran here.
@noble/post-quantum build; if that module cannot load, the page falls
back to a SHA-256 hash chain and labels every receipt accordingly, so it never presents an unverified
signature as verified. USPTO patent portfolio pending. Contact [email protected].