CABIER AI · Global assurance
Global AI Assurance
Governance asks who owns AI. Assurance asks whether the institution can prove its policies operate. As systems retain memory, use tools, reach data and take actions, the supervisory question moves from what was approved before deployment to what the estate is doing right now, and who is able to stop it.
CABIER is model-agnostic by design. The institution chooses the models. CABIER assures the estate that results, including the parts of it that were never procured as AI.
Published as reference architecture. Control weights, grading rubrics and the full control set are set per engagement.
12
AI assurance modules
320
AI control templates
24
Mapped jurisdictions
8
Trust Gates
Figures as at September 2026.
See Cabier at work: run the AI scenarioThree assurance subjects, one AssureCore boundary
The model itself, the system that deploys it, and the regulatory regime that governs it. Jurisdiction and sovereignty cut vertically through all three, so a frontier model and a deployed agent are assessed on identical terms.
Model assurance
Is the capability understood, evaluated and appropriately controlled?
System assurance
Is the deployed AI operating within authorised boundaries?
Regulatory assurance
Can supervisors define, monitor and supervise AI consistently across jurisdictions?
Cutting vertically through all three
Sovereignty is never read from where a server sits. Applicable requirements come from the combined relationship between entity, activity, data, user, model, infrastructure, transaction, jurisdiction, sector and consequence.
Cutting horizontally through all three
Systemic assurance
A model can be safe, an institution can be compliant, an agent can be authorised, and the ecosystem can still become unsafe through interaction effects. Systemic assurance is the layer that asks what happens when millions of AI-enabled decisions interact at once across an interconnected system.
Cabier does not predict financial crises and does not claim to. It identifies emerging deviations, correlations, concentrations and propagation patterns that may indicate systemic stress or coordinated disruption, states its confidence, shows its derivation and escalates to human authority.
How a consequential action moves through the architecture
One spine, read top to bottom, with the intelligence foundation underneath it. Each named layer is a surface you can open.
- 01
Frontier, enterprise and sovereign AI
The systems entering the environment, whatever their provenance.
- 02
AssureAdapt
Consequence-driven assurance routing. Sets both assurance intensity and computational intensity.
- 03
Deterministic controls, TrustGraph and Assurance Intelligence
Three parallel inputs to the decision, so no single one is a required dependency.
- 04
AssureCore
The eight Trust Gates and the disposition, with governing facts, authority, controls, evidence and escalation condition.
- 05
AssureMark
Evidence and provenance, written as machine-readable assurance context with an expiry.
- 06
Human and institutional authority
Where consequential decisions are taken. Assurance prepares them, it does not replace them.
- 07
Assurance state
The current, dated, derivable position, and the movement in the institution's resilience position.
Foundation
Cabier Assurance Corpus
Horizontal ontology and vertical domains, with relationships preserved.
Cabier Assurance Intelligence
Developed and evaluated against the corpus. Model-independent and model-diverse.
Continuous learning, evaluation and reassessment
Governed pipeline, versioned, with human-reviewed outcomes only.
Intelligence + authority + context + consequence = assurance
The machine asks the same nine questions of every consequential action, whether the actor is a person, a model or an agent instructed by another agent.
The architecture, and the order it is built in
Published so a reader can see the whole system and the sequence. Only surfaces marked live are available today; everything else is published as architecture.
The loop
Bounded throughout by
Build sequence
- Phase One
AssureCore, TrustGraph and the Universal Assurance Seed as the common vocabulary. - Phase Two
Frontier model assurance, AssureMark and AssureAdapt. - Phase Three
Agent assurance and the embedded API and gateway surfaces. - Phase Four
Sentinel and adversarial intelligence. - Phase Five
Systemic intelligence and financial flow intelligence. - Phase Six
Regulatory and sovereignty intelligence at depth. - Phase Seven
Vertical assurance packs over one architecture. - Phase Eight
Developer ecosystem and Cabier Assured.
A sequence is not a delivery commitment and no date is implied.
Nine disciplines under one umbrella
Each discipline exists because a different person asks the question. The board, the regulator, the CISO, the auditor and the operator are not asking the same thing about the same system.
AI Governance
Who owns AI, and under what policy?
Named accountability per system, an approved policy set and a board record of the risk that was accepted.
AI Assurance
Can the institution prove those policies operate?
Evidence produced continuously by the control itself, not assembled when a supervisor asks.
AI Control
Can the institution constrain what AI is able to do?
Scoped authority per agent, enforced at the point of action rather than described in a policy document.
AI Security
Can it detect and contain an AI-enabled attack?
Prompt injection, jailbreak, privilege escalation and exfiltration treated as cyber events with owners.
AI Resilience
Can it survive model, provider or agent failure?
Fallback models, exit tests and impact tolerance measured with the AI dependency as the failure mode.
AI Regulatory Intelligence
Which rules actually apply, and from when?
Obligation-level mapping from authority to control to evidence, with effective dates and supervisory expectation.
AI Sovereignty
Where do the model, the data and the evidence operate?
Residency and compellability resolved per workload before deployment, not discovered during an examination.
AI Supply Chain
Which third-party models, agents and tools are embedded?
The estate the institution did not choose but already owns through a vendor, mapped to the dependency beneath it.
AI Agent Assurance
What can autonomous systems actually do today?
Authority envelope per agent, an out-of-envelope refusal on record and a named human owner for each.
The architecture
Models and agents sit below. The assurance layer sits above them, and everything it produces resolves into one institutional position.
Global model ecosystem
Ecosystems observed in institutional estates. Naming is not endorsement, and no comparative benchmark is published.
CABIER assurance layers
AssureCore
Eight Trust Gates resolve identity, authorisation, policy, data, security, regulatory, resilience and evidence conditions.
TrustGraph
One typed map of obligations, controls, systems, models, agents and evidence.
Decision Fabric
Where an intervention is decided, recorded and attributed.
Evidence Vault
The artefact a supervisor, auditor or board reads.
AI Trust Score
Per-system position across fourteen dimensions.
ORS
The institutional resilience position the board is accountable for.
Enterprise
Board, CRO, CISO, CCO, CIO and internal audit read the same position from the same evidence.
Government
Ministries, agencies, supervisors and auditors read it under their own residency conditions.
One assurance foundation, composed per institution
AssureCore and TrustGraph, plus a jurisdiction pack, an industry pack and the enterprise AI estate. Different obligations and evidence, with the same decision boundary underneath.
| Jurisdiction | Industry | Pack | What changes |
|---|---|---|---|
| Canada | Banking | OSFI E-23 and B-10 with federal privacy overlay | Model risk rating drives validation depth, and third-party AI is examined at the same time. |
| United States | Insurance | NAIC model bulletin with state-level conduct overlay | Conduct and unfair-discrimination testing sit ahead of model documentation. |
| United Kingdom | Legal | SRA and professional-privilege constraints with SMCR-style accountability | Privilege and confidentiality bound the reach of the agent before any performance question. |
| Singapore | Government | Sovereign deployment with public-sector data classification | Residency and compellability decide the architecture, not the procurement. |
Embedded and infrastructure assurance
The same primitives, distributed through the stack
Assurance that sits in a separate system describes the past. These surfaces place the same identity, authority, control, evidence and human authority sequence where intelligence is manufactured, hosted, delegated and acted upon. Published as reference architecture.
01AI factory
Where intelligence is manufactured: accelerators, clusters, training and the model estate itself.
02Data centre, cloud or sovereign infrastructure
Where intelligence is operated, by a named operator, under a legal relationship.
03Model
The capability, at a stated version, with evaluation evidence that belongs to that version.
04Runtime
Where inference and agent execution actually occur, rather than where they are described.
05Agent
The acting party, with an objective, a delegation path and an authority envelope.
06Tools, tool servers and credentials
The reach an agent acquires the moment a tool or a credential is attached.
07Data
Classification, residency, permitted purpose and onward flow of everything reachable.
08Application
Where intelligence is distributed to people and to other systems.
09Enterprise and device
Where intelligence is consumed, inside an institution or at the edge.
10Action
The consequential thing that happens, which is the only place assurance finally matters.
CABIER spans every layer
TrustGraph
The relationships between every object in the hierarchy, so an effect can be traced rather than guessed.
AssureAdapt
Detects change anywhere in the hierarchy and decides when the position must be re-derived.
AssureCore
The single decision boundary: eight Trust Gates, five dispositions, no consequential autonomous execution.
AssureMark
The machine-readable assurance and provenance record for whatever was decided.
Evidence
One evidence architecture, dated, attributable and re-examinable by someone who was not there.
Human and institutional authority
The named authority that adjudicates consequential findings, at every layer.
Cabier does not replace the underlying infrastructure. Cabier assures what operates across it.
Capability
Frontier Model Assurance
Provider evidence is not the same thing as independently verifiable assurance, and the difference is recorded.
Capability
Independent Evaluation
Evaluation under a stated protocol, scope and consequence class, with limitations recorded alongside results.
Reference architecture
Frontier Model Assurance Exchange
Reciprocal and independent evaluation across eighteen domains. Evaluation is evidence, not assurance.
Control plane
AI Control Plane
Model and version governance, ownership, residency, approval status and enterprise tool governance.
Control plane
Agent Assurance
Identity, permission fabric, runtime assurance and security framework for agents that act.
Reference architecture
Agent Lineage
Identity, objective, authority, delegation, action and intervention resolved across the existing agent surfaces.
Capability
Runtime Assurance
Behaviour where it actually occurs, rather than where it was described at approval.
Reference architecture
Embedded Assurance
Above, beside, within the boundary, at the gateway or in the runtime, with compute able to stay with the customer.
Reference architecture
AI Factory Assurance
Compute, clusters, model estate, routing, inference, agents, credentials and tools as one governed environment.
Reference architecture
Data Centre Assurance
The site as an assurance subject, where a residency statement is either evidenced or merely asserted.
Assurance surface
AI Data & IP Assurance
Where did the data go, under whose authority, and was training use permitted.
Cross-cutting analysis
Behavioural Selection Risk
Where individually approved components combine into an outcome nobody approved.
Cross-cutting lens
Systemic Assurance
Concentration, dependency and cross-system interaction read together, not as a fifth vertical.
As intelligence becomes more capable and more autonomous, assurance becomes infrastructure. That is the layer CABIER operates.
Start with estate discoveryFlagship analysis
Governance became an operating control
The full argument: why the approval gate stopped working, what a running assurance loop does instead, and what the board should be told rather than scored.
Read the article