You cannot govern an estate you have not counted.
Nearly every AI governance failure we review starts the same way: the institution did not know how many systems it was running, who owned them, or which jurisdiction's rules applied. The inventory is the first control, and every other AI control depends on it.
You choose the AI. Cabier orchestrates the trust.
Intake and lifecycle
One declaration produces every jurisdictional classification the institution owes.
Declaration
Every AI system entering the estate is declared with purpose, owner, jurisdiction of use and data classes touched.
Classification
Risk tier assigned per jurisdiction — EU AI Act category, SR 11-7 model tier, OSFI E-23 rating — from one declaration.
Validation
Independent review before production. Self-assessment is not accepted as validation evidence.
Approval
A named executive signs the deployment. The signature and its scope are recorded, not implied.
Monitoring
Drift, eval regression, incident rate and evidence completeness tracked continuously against the approved scope.
Retirement
Decommissioning is a controlled event with evidence retention obligations preserved.
Inventory explorer — published sample
Search and filter an illustrative estate of eight systems. Select a record to see its accountable owner, residency, classification and the evidence bound to it.
Risk tier
Status
Record · AIS-0114
Credit adjudication assistant
- Accountable owner
- Chief Credit Officer
- Model family
- Frontier general-purpose model, vendor-hosted
- Hosting and residency
- EU region, single-tenant inference
- Jurisdictions of use
- EU, UK
- Classification
- High risk · In production
- Last independent validation
- 12 May 2026
Evidence bound to this record
- Evaluation resultsEV-24118 · v3
Credit-memo faithfulness eval, 480-file holdout set
Control MLC-04 · captured 12 May 2026 · reviewed by Independent validation
- Human oversight recordsEV-24120 · v11
Adjudicator amendment log with sign-off scope
Control AIG-03 · captured Continuous · reviewed by Chief Credit Officer
Illustrative, non-client sample. Evidence specifications, effectiveness grading and the full inventory schema are released under engagement.
Evidence viewer
Six evidence classes bound to the control they satisfy, versioned and time-aware.
Benchmark and task-specific eval runs with version, dataset and date. Reruns are versioned, never overwritten.
Protected-attribute analysis where the use case makes it applicable, with methodology recorded.
Adversarial and prompt-injection results, severity, and the remediation that closed each finding.
Citation fidelity and grounded-answer rates for retrieval-backed systems.
Where a human reviewed, what they saw, what they changed, and whether they signed.
Supplier declarations for third-party models, treated as claims requiring confirmation.
AI incident register
AI failures are operational incidents. They run through the same register, the same materiality logic and the same supervisory clocks as everything else.
How the failure surfaced — monitoring, human report, downstream control or customer complaint.
Assessed against the same materiality logic used for cyber and operational incidents.
Whether the system was quarantined, restricted or left running under supervision, and who decided.
Which supervisory clocks started, in which jurisdictions, and when each was met.
Model, data, prompt, integration, oversight failure or scope creep — categorised, not narrated.
Control change, evidence refresh and score impact recorded to the vault.
The inventory feeds the AI Trust Score, which feeds the Operational Resilience Score. No dimension is scored from a self-assessed status.