Governance in the state your AI acts on.
Your agent calls a tool, gets an answer back, and acts on it. Nothing in that answer says where it came from or whether the agent was allowed to use it. Rootz attaches both, so the agent can check before it commits you to anything. An output that carries the conditions it was produced under is what we call Measured AI.
Governance is part of agentic state.
An agent’s next action depends on the state it carries: its authority, the inputs it has accepted and the conditions its policy requires. Rootz makes the supporting evidence part of that state as work moves between systems.
The organization defines policy. Verification tools check the messages and evidence before the agent proceeds. The result becomes part of the next exchange, with an exception path when the required checks fail.
The evidence behind an answer
Inputs with Origin
Retain the source, version and retrieval context of the data used. Distinguish original records from derived extracts and summaries.
Measured data quality
Specify the tests the task needs: required fields, coverage, age, consistency or numerical checks. Preserve the method and result, rather than collapsing quality into one score.
Authorized instructions
Bind the task to the requesting principal, its scope and the intended recipient. Validate authority separately from signature correctness.
Processing evidence
Attach the measurements available for the deployed service. Model weights, application code, boot state and facility assertions have different evidence requirements.
Checkable results
Return a signed result with its request binding and evidence. A receiving agent can accept it, request stronger evidence or route an exception under customer policy.
The next system receives the evidence too.
A result becomes an input to another agent, API or government application. Its signed request binding and evidence allow the recipient to apply its own policy. Governance continues across that exchange.
Follow the request and resultTwo useful starting points
An evidence-aware data assistant
Connect one prepared source to one agent. Define freshness and coverage requirements, then demonstrate how a missing or altered record changes the agent’s next step.
A signed inference or tool service
Connect one request to one response and the available process evidence. Test whether the consumer can detect an altered answer, a mismatched request and an expired record.
The security direction is already explicit.
The April 2026 joint Five Eyes guidance recommends signed authorized instructions and fresh cryptographic proofs for privileged calls. The NSA’s May MCP guidance recommends message signatures, expiry, replay protection and binding to time and context. Rootz implements mechanisms in this direction; each deployment documents its actual coverage.
Make one AI exchange verifiable.
Bring the service, the consuming agent and the person who owns its acceptance policy. We will define the evidence contract with your team.
Scope the integration