A defensible score for every AI system in your estate.
Fourteen evidence-anchored dimensions produce one institution-specific score per AI system. The output is a live band — Institutional, Supervised, Restricted or Quarantine — feeding the same Operational Resilience Score the board and regulator already read.
You choose the AI. Cabier orchestrates the trust.
The ten dimensions
Categories, not weights.
The rubric is published as fourteen categories. Exact weightings, thresholds and scoring internals are institutional IP and stay off public surfaces.
Security
Threat model coverage, red-team results, supply-chain posture.
Reliability
Uptime, drift, regression rate under production load.
Hallucination
Grounded-answer rate, citation fidelity, factuality tests.
Governance
Registry entry, approvals, accountable owner, review cadence.
Compliance
Jurisdictional obligations mapped and evidenced (EU AI Act, SR 11-7, NIST, ISO 42001).
Resilience
Failover, graceful degradation, dependency mapping, incident history.
Cyber posture
Prompt-injection defence, secrets isolation, egress control, identity.
Explainability
Reasoning trace, feature-level interpretability where applicable.
Evidence completeness
Prompts, outputs, decisions, overrides recorded to the vault.
Operational maturity
Change control, rollback, monitoring, runbooks, human sign-off.
Agent authority
Declared authority envelope, ceilings honoured, out-of-envelope refusals on record.
Runtime behaviour
Live action stream against the approved envelope, including intervention history.
Incident history
Structured AI incident record, control that failed, and what changed afterwards.
Sovereignty
Residency of model, inference, logs and evidence, and which authority can compel access.
Score bands
Four bands. One control action.
Institutional
Approved for regulated business functions without additional supervision.
Supervised
Approved with human-in-the-loop and elevated monitoring.
Restricted
Limited to non-material use cases; remediation plan required.
Quarantine
Not approved for production. Registry-locked pending review.
Sample card (illustrative)
How a card reads.
Registry entry
Frontier chat model · v-family · fine-tune F-12
Band
Supervised
Security
Evidence on file
Reliability
Evidence on file
Hallucination
Evidence on file
Governance
Evidence on file
Compliance
Evidence on file
Resilience
Evidence on file
Cyber
Evidence on file
Explainability
Evidence on file
Evidence
Evidence on file
Ops maturity
Evidence on file
Human-in-the-loop required · registry approval expires 90 days
Illustrative only. No live vendor is publicly scored on Cabier surfaces.
Guardrails