EVE AI Core
EVE AI Core
Borrowed from networking, a control plane is the layer that decides; the data plane is the layer that does. An AI control plane is where every proposed AI action is checked against policy before it runs.
“An AI control plane is a dedicated layer between an AI model and the actions it can take that decides — before execution — whether each proposed action is allowed under policy.”
The term comes from networking. A router’s data plane forwards packets; its control plane decides where packets are allowed to go. AI borrows the split: the model and application are the data plane that produces outputs and takes actions; the control plane is the separate layer that decides what those actions may be — enforced the same way no matter which model or prompt produced the request.
An application or agent does not act directly. It proposes an action; the control plane evaluates it against policy and returns a verdict; only an allowed action executes — and every outcome is recorded.
The application hands the control plane a proposed decision or action instead of performing it. Nothing has happened yet — which is exactly why a “no” is possible.
The control plane checks the proposal against the active policy — eligibility, prohibited factors, approval limits, hard safety rules — and returns allowed, modified, or blocked.
Only an allowed action runs. Every outcome — including blocks — emits a signed record that can be verified later, independent of the application’s own logs.
A control plane is defined by three properties: it sits in front of the action, not after the output; its verdict is binding, not advisory; and it produces evidence of every decision. A layer missing any of those is a filter, not a control plane. EVE’s implementation is deterministic — the same input and policy version always yield the same verdict — which is what makes its decisions reproducible and auditable. That is the subject of the deterministic AI governance control plane.
The control plane is the where. The next questions are how its decisions are made reproducible, and how that beats a probabilistic filter.
A dedicated layer between an AI model and the actions it can take that decides, before execution, whether each proposed action is allowed under policy.
A guardrail scores or filters output after generation and the output may still ship; a control plane returns a binding verdict before the action runs and records signed evidence.
Because the decision and its consequence are separate events — checking first means a blocked decision never reaches production.
See how a deterministic control plane intercepts a real decision, returns a binding verdict before execution, and signs the evidence.