LAB-01
AI Control Plane
A semantic control layer that makes every AI decision traceable, policy-enforced and auditable, without changing the code that calls the model.
- Status
- parked
- Note
- The deployment sat on a third-party host whose subscription lapsed, so the old address serves a hosting-provider notice instead of the product. It has been removed rather than left up, and there is nowhere to send a visitor until the system is self-hosted.
- Stack
- Policy-as-code · Knowledge graphs · Audit trails · LLM gateways
Known issues
- There is no public address for this system. The previous one served a hosting-provider notice rather than the product, so it was removed instead of left in place.
- The public address depended on a third-party hosting subscription rather than on infrastructure Clemence W. Chee controls, which is what allowed a lapsed membership to take the site down.
- Self-hosting is the fix and it has not been scheduled, so the entry stays parked until it is.
Clemence W. Chee built the AI Control Plane, or AICP, as a semantic governance layer that sits between an application and the language model it calls.
What does the AI Control Plane do?
AICP records every AI decision as an immutable artefact carrying the context that produced it.
The gateway sits in the request path. An application calls AICP instead of calling the provider directly, and AICP forwards the call, applies the policies that govern it, and writes the decision to an audit trail that cannot be edited afterwards. Integration is a one-line change at the call site.
Why does it exist?
Organisations deploy language models faster than they build the supervision for them.
The gap is specific. A model in production makes decisions that affect customers. Most organisations cannot answer three questions about any single one of those decisions: which data produced it, which policy allowed it, and who is accountable for it. AICP answers all three as running infrastructure.
What is inside it?
Three parts: an immutable audit trail, a metadata knowledge graph, and a runtime policy engine.
- Audit trail. Every decision is written as a tamper-evident artefact with the prompt, the context, the model version and the outcome attached.
- Metadata knowledge graph. Datasets, features, models, prompts, decisions and outcomes are connected as nodes, so provenance is a graph traversal.
- Runtime policy enforcement. Governance rules are written in YAML and evaluated before execution, so a call that violates a policy never reaches the provider.
Which compliance regimes does it target?
AICP is designed to support evidence collection for SOC 2, GDPR, PCI-DSS, HIPAA and FDA 21 CFR Part 11.
That is a design target and a claim about the artefacts the system produces. It is not a certification, and nothing on this page should be read as one. An auditor still audits.
What is the current status?
The architecture described above is real and the deployment is down. It ran on a third-party host whose subscription lapsed, so the old address serves a provider notice.
The public address has been removed rather than left pointing at a notice. The February 2026 essays on AI control planes describe the same architecture, and that description is what this page can currently stand behind. Clemence W. Chee is self-hosting the deployment, and this entry stays parked, with no address on it, until that is running.