Cabier Global Assurance · Architecture

    Assurance Corpus and Assurance Intelligence

    An assurance system that reasons only by calling a general-purpose model has a structural problem: it depends on the same class of intelligence it is meant to assure. The foundation layer exists to remove that dependency.

    The Cabier Assurance Corpus is the structured knowledge foundation. Cabier Assurance Intelligence is a specialised reasoning layer developed and evaluated against it, and it is deliberately model-independent.

    Published as reference architecture. The corpus contents, evaluation methodology and reasoning implementation are proprietary and are not published.

    Horizontal ontology

    Twenty-four object classes that recur in every regulated and high-consequence setting, whatever the sector calls them.

    EntityIdentityAuthorityActivityAI and modelAgentDataSystemToolTransactionDependencyJurisdictionLawRegulationStandardObligationPolicyControlEvaluationEvidenceIncidentRiskConsequenceOutcome

    Vertical overlays

    The sector changes the obligations and the language. It does not change the object model.

    Financial services

    Prudential, conduct, financial crime, payments, resilience and market integrity obligations.

    Insurance

    Underwriting, pricing, claims conduct, capital and reserving.

    Legal

    Privilege, confidentiality, conflicts and professional duty.

    Healthcare

    Clinical safety, patient data, device and decision-support duties.

    Government and public sector

    Public accountability, procurement, records and national residency expectations.

    Critical infrastructure

    Operational technology, continuity, incident duties and national resilience.

    Digital assets

    Custody, settlement, reserve assurance and tokenised instrument obligations.

    Cybersecurity

    Access, detection, response, reporting clocks and third-party dependency.

    Energy

    Reliability, safety, environmental duty and dispatch integrity.

    Industrial and logistics

    Safety, provenance and physical-system authority.

    Social and algorithmic platforms

    Systemic risk assessment, content and recommender accountability, minors, transparency and researcher access.

    Relationships, not documents

    The corpus preserves relationships rather than storing source material as disconnected documents. A single chain has to read end to end: law, obligation, entity, system, model, agent, authority, control, evidence, outcome. That is what makes a conclusion derivable rather than asserted.

    LawObligationEntitySystemModelAgentAuthorityControlEvidenceOutcome

    What the intelligence layer reasons about

    Authority and permission reasoningObligation resolution across simultaneous regimesControl derivation and control adequacyEvidence sufficiency and freshnessRisk, dependency and propagation reasoningJurisdiction and sovereignty reasoningAI behaviour and agent authority reasoningConsequence and resilience reasoning
    • Cabier Assurance Intelligence is a specialised assurance reasoning layer. It is not a general-purpose frontier model and does not compete with frontier providers on general intelligence.
    • It is model-independent by design. It may draw on Cabier models, open-weight models, customer-controlled models, approved frontier models, deterministic reasoning, statistical methods and graph reasoning.
    • No single external model is a mandatory dependency. Cabier must never depend exclusively on the intelligence provider it is assuring.
    • It is developed and evaluated against the Cabier Assurance Corpus, leaving the implementation route open across retrieval, graph reasoning, specialised models, distillation, human feedback and synthetic evaluation.

    Assurance intelligence independence

    An assurance decision must be capable of being produced through deterministic controls, TrustGraph, the Assurance Corpus, Cabier Assurance Intelligence, independent model evaluation and human authority, rather than requiring any one frontier model. That is the answer to who assures the assurer.

    Deterministic controls

    The mechanical part of an assurance decision resolves without inference at all.

    TrustGraph

    Relationships, dependencies and propagation are read from the graph, not guessed.

    Assurance Corpus

    Structured authority, obligation, control and evidence knowledge with preserved relationships.

    Cabier Assurance Intelligence

    Specialised reasoning where the answer is a judgement rather than a lookup.

    Independent model evaluation

    More than one assessment for high-consequence questions, so no single model is the arbiter.

    Human authority

    A named person or committee holds the consequential decision. Assurance prepares it.

    The assurance learning cycle

    Eight steps, with the important constraint stated immediately after them.

    01

    Observe

    Behaviour, permissions, obligations, dependencies and evidence freshness.

    02

    Evaluate

    Against the control set and the evaluation requirements for that consequence class.

    03

    Decide

    Through the Trust Gates, with the disposition stating its basis.

    04

    Review

    Human review where the consequence class or the ambiguity requires it.

    05

    Evidence

    The decision, its basis and its outcome recorded and dated.

    06

    Outcome

    What actually happened, including refusals, interventions and near misses.

    07

    Learn

    Reviewed outcomes improve the corpus and subsequent evaluation through a governed pipeline.

    08

    Reassess

    The assurance position is recomputed rather than assumed to still hold.

    • Production decisions do not automatically become training data.
    • Material entering the corpus passes a governed pipeline: provenance, approval, privacy controls and versioning.
    • Customer and institutional material is used only within the permissions granted for it, and residency constraints travel with it.
    • Every corpus version is identified, so a past conclusion can be re-derived against the knowledge that existed at the time.

    The spine

    How a consequential action moves through the architecture, and what sits underneath it.

    1. 01

      Frontier, enterprise and sovereign AI

      The systems entering the environment, whatever their provenance.

    2. 02

      AssureAdapt

      Consequence-driven assurance routing. Sets both assurance intensity and computational intensity.

    3. 03

      Deterministic controls, TrustGraph and Assurance Intelligence

      Three parallel inputs to the decision, so no single one is a required dependency.

    4. 04

      AssureCore

      The eight Trust Gates and the disposition, with governing facts, authority, controls, evidence and escalation condition.

    5. 05

      AssureMark

      Evidence and provenance, written as machine-readable assurance context with an expiry.

    6. 06

      Human and institutional authority

      Where consequential decisions are taken. Assurance prepares them, it does not replace them.

    7. 07

      Assurance state

      The current, dated, derivable position, and the movement in the institution's resilience position.

    Foundation

    1. Cabier Assurance Corpus

      Horizontal ontology and vertical domains, with relationships preserved.

    2. Cabier Assurance Intelligence

      Developed and evaluated against the corpus. Model-independent and model-diverse.

    3. Continuous learning, evaluation and reassessment

      Governed pipeline, versioned, with human-reviewed outcomes only.

    Assurance of the assurance system

    AI is assured by Cabier. Cabier's own assurance mechanism is assured through evidence, testing, governance and human authority. The recursion is manageable because it terminates in the same place every institutional control does: a named human accountable for a dated decision.

    Model and version provenanceControl-version provenanceCorpus-version provenanceEvidence provenanceEvaluation of the assurance mechanism itselfDrift monitoringBias and error testing where relevantHuman override trackingReproducibility of a past conclusionIndependent testingAudit trailChange management

    Where the same reasoning is applied to interaction effects across an interconnected ecosystem.

    See the systemic layer