GateHouse
All papers
Architecture

The metadata-only control plane

24 pagesJul 2026

Abstract

GateHouse runs a commercial network across sovereign environments without ever centralising raw data. This paper details the control-plane / data-plane split, the recursive disclosure gate, and a precise definition of what “metadata-only” means — including for GateHouse’s own telemetry and its in-container AI.


The sovereignty problem

A network needs a shared view to function — discovery, matching, benchmarking all require some common surface. But the whole value proposition of GateHouse is that raw data never leaves its owner. The architecture exists to reconcile those two facts: a functioning network whose control plane never holds the asset it commercialises.

Control plane / data plane

GateHouse adopts the well-understood control-plane / data-plane split. The data plane runs inside each participant’s environment as a container; it is where profiling, policy evaluation, workload admission and metering happen. The control plane is metadata-only: it coordinates, it never computes on raw data, and it never receives it.

Sovereignty here is architectural, not contractual. The claim is not “we promise not to look” — it is “the raw data is never sent.”

The disclosure gate

Even metadata is disclosure. So metadata must pass the owner’s disclosure policy before it leaves the container — including telemetry destined for GateHouse itself. The gate is a single primitive: a policy-evaluated checkpoint that decides what may cross a given boundary, for a given audience, under given attributes.

One gate, many boundaries

The same gate primitive is applied recursively. It governs what leaves the organisation to the network, what reaches a specific counterparty, what reaches GateHouse, and — where an organisation chooses tighter internal controls — what flows from an individual container up to the organisation’s own dashboards. Any surface renders exactly what has passed the gates between it and the data.

recursive gate — container → org → network → GateHouse
The disclosure gate applied at every boundary. Each surface sees only what passed the gates beneath it.

In-environment AI

In-container agents default to an in-environment LLM, so prompts, intermediate reasoning and staged outputs never leave the boundary. Where a model-IP-protected trust branch is required, confidential-computing substrates (TEEs, remote attestation) are available — the architecture is designed to accommodate that branch without redefining the gate.

What crosses, precisely

Three classes of object cross a GateHouse boundary, and only these: a permission granted under policy; a model or workload admitted to run in place; and a computed insight that has passed the gate. Raw records are not on the list, and the boundary counter that proves it reads zero.

Evaluate it against your requirements.

We will map the architecture and governance model to your estate, jurisdiction and constraints.