Supervisor middleware example for an external consistency verifier. #4052
Anaciber
started this conversation in
Vouch Request
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
What do you want to work on?
I'm Ignacio Ozcáriz, CEO of Data Farm Network (Spain), a company belonging to the Criptosasun Energia group that is a member of NVIDIA Inception. Our work consists of the certification of consistency between Artificial Intelligence agents: the idea is to verify, outside the reach of the agents themselves, that several of them have not consumed the same versioned resource twice. We want to provide two things: an example under examples/ with a middleware supervisor that consults an external verifier and denies the second consumption of a resource before the request reaches the resource, with a stub verifier and a contract test; and a line in the extensibility documentation on the openshell.supervisor-middleware.contract capability, which today is not documented and causes the middleware registry to fail with an unexplanatory error. We already have it working end-to-end in OpenShell 0.1.2, on a Raspberry Pi, with two sandboxes under the same policy. I'll start with a feature request and, if the maintainers deem it necessary, I'll propose it as an RFC.
Why this change?
OpenShell guarantees the perimeter of each agent and traces its actions, but unfortunately now nothing on the platform says if the joint result of several agents was consistent: two agents in different sandboxes can consume the same version of a resource and both come out fine. Today that gap is closed by putting a person at the end of the chain or analyzing the logs later, and neither of the two prevents the second action. The middleware supervisor already sees each request before the credentials and can deny it; what is missing is a convention to describe a request as an action on a shared resource and to record the verdict. Our work comes as the culmination of the implementation in the environment of distributed information processing, which in the end is the definition of the ecosystem of AI Agents, of the conclusions of our paper "The Lost Update as a Double Spend: Validated Ordering and Global Consistency in Distributed Data Systems“ (https://ssrn.com/abstract=7345423) . We believe that this development can be very useful for anyone who deploys agents on orders, payments, quotas or registrations, and that is why we want to contribute to it openly.
Checklist
All reactions