AI Assurance · Control plane
Cabier AI Control Plane
A control architecture, not a model. The institution stays model-agnostic; the plane decides which model may be used for which requirement, under whose authority, with what evidence produced. Every path is human-governed — nothing here executes without an accountable owner.
Published as reference architecture. Deployment scope, interoperability and residency options are set per engagement.
Model registry
Extends the existing AI model inventory rather than duplicating it. Each field exists because a supervisor, an auditor or a resilience test asks for it.
Model and version
The unit of supervisory scope; version is what is examined, not the family.
Accountable owner
A named human, distinct from the validator.
Hosting region and residency
Determines transfer mechanism and sovereign constraints.
Approval status
Blocks production use where validation has not closed.
Security rating
Adversarial and prompt-injection posture, refreshed on model change.
Bias and explainability status
Feeds fairness obligations across FEAT, Art. 10 and fair-lending.
Cost and latency envelope
Routing constraint and a resilience input, not a finance metric.
Last validation and next due
Drives revalidation cadence and supervisory readiness.
Evidence chain reference
Every field resolves to an artefact in the vault.
Model router
Requirement in, model out, with the reason recorded and an audit identifier attached. Routing is a control decision before it is an engineering one.
| Requirement | Routed to | Reason |
|---|---|---|
| Personal data of EU residents, high-risk classification | EU-hosted model, no egress | Chapter V transfer constraint plus Art. 10 data governance. |
| Sovereign citizen data, Gulf deployment | In-country inference endpoint | National charter prohibits egress for citizen-service systems. |
| Broker-dealer client communications | Retained, supervised, archived model path | FINRA 3110 supervision and books-and-records retention. |
| Internal research summarisation, no personal data | Lowest-cost qualified model | No regulated activity engaged; cost and latency govern. |
| Credit decisioning | Validated, explainable model only | SR 11-7 plus fair-lending explainability obligations. |
Enterprise tool governance
Reach
Which systems an agent may contact, declared before it runs, not discovered afterwards.
Conditions
The state the institution must be in for the action to be permitted at all.
Ceilings
Value, volume and blast-radius limits that stop the agent rather than warn about it.
Log
Every reach, permitted or refused, written to the evidence chain with an audit identifier.
Multi-model adapter grid
Frontier commercial
Contractual right-to-test, sub-processor disclosure, exit path
Open-weight self-hosted
Provenance attestation, in-house adversarial testing, patch cadence
Sovereign or regional
Residency proof, in-country inference, national charter mapping
Embedded vendor model
Discovery first — the model the institution did not know it bought
The plane is the control layer above whichever models an institution chooses — and above the ones it did not choose but already owns through a vendor.
Readiness dimensions