Certification & Trust Tiers

CAR defines two distinct tier systems, each with 8 levels (T0–T7), that work together to determine an agent's effective autonomy.

Dual-Tier Architecture

System	Scope	Changes When	Determines
Certification Tier	External attestation status	Re-certification, audit	What agent could do
Runtime Tier	Deployment-specific autonomy	Continuously (per-action)	What agent may do now

Effective Autonomy = MIN(Certification_Tier, Runtime_Tier)

A T6-certified agent deployed in a T3-rated context operates at T3.

Certification Tiers

External, point-in-time assessment by a certification authority.

Tier	Name	Score	Color	Description
T0	Sandbox	0–199	Stone	No external verification, testing only
T1	Observed	200–349	Red	Identity registered, behavior monitored
T2	Provisional	350–499	Orange	Initial capabilities verified
T3	Monitored	500–649	Yellow	Continuous monitoring active
T4	Standard	650–799	Green	Standard certification achieved
T5	Trusted	800–875	Blue	Full trust established
T6	Certified	876–950	Purple	Third-party audit completed
T7	Autonomous	951–1000	Cyan	Highest assurance level

Runtime Tiers

Continuous, context-aware trust evaluation by the Vorion Trust Engine.

Same names and score ranges as Certification Tiers, but computed dynamically:

Runtime Score = (Certification × 0.3) + (Behavior History × 0.4) + (Context × 0.3)

Scoring Components

Component	Weight	Sources
Certification	30%	CAR certification tier, attestation status
Behavior History	40%	Success rate, policy violations, drift events
Context	30%	Deployment environment, regulatory ceiling, org policy

Tier Transitions

Score ↑ (good behavior over time) → Tier promotion
Score ↓ (violations, drift, anomaly) → Tier demotion
Revocation event → Immediate drop to T0

Regulatory Ceilings

Compliance frameworks impose maximum trust scores:

Framework	Max Score	Max Tier	Retention
EU AI Act	699	T4	7 years
ISO 42001	799	T4	5 years
NIST AI RMF	899	T5	5 years

An agent in an EU AI Act-regulated context can never exceed T4 regardless of its certification.

Tier-Based Permissions

Tier	Max Level	Delegation	Key Privileges
T0–T1	L2	None	Read-only + advisory
T2	L3	None	Supervised execution
T3	L4	Depth ≤ 2	Autonomous within bounds
T4	L5	Depth ≤ 2	Expanded scope
T5–T6	L6	Depth ≤ 4	Agent spawning
T7	L7	Depth ≤ 8	Sovereign authority

Trust Score API

const client = new CARClient({ endpoint: 'https://api.agentanchor.io' });

// Get current trust evaluation
const trust = await client.getTrustScore('a3i.vorion.banquet-advisor:FHC-L3@1.2.0');

console.log(trust);
// {
//   score: 742,
//   tier: 4,
//   tierName: 'Standard',
//   components: {
//     certification: 0.8,    // T5 certified
//     behavior: 0.72,        // Good history
//     context: 0.65          // Standard deployment
//   },
//   ceiling: { framework: 'NIST', maxScore: 899 },
//   evaluatedAt: '2026-02-08T12:00:00Z'
// }

Gaming Detection

The CAR specification includes anti-gaming measures:

• Sudden score jumps trigger manual review
• Oscillating behavior (good/bad cycles) limits maximum achievable tier
• Context switching (rapid environment changes) resets behavior scoring
• Provenance modifiers affect starting trust for cloned/imported agents