CAR Architecture Overview

The CAR system is built on a three-layer architecture that separates identity, certification, and runtime governance concerns.

Layer Model

┌──────────────────────────────────────────────────────────┐
│  Layer 3: Semantic Governance & Runtime Assurance         │
│  Behavioral monitoring, drift detection, policy engine    │
├──────────────────────────────────────────────────────────┤
│  Layer 2: Capability Certification & Extensions           │
│  Attestations, extensions, trust scoring                  │
├──────────────────────────────────────────────────────────┤
│  Layer 1: Identity & Trust Primitives                     │
│  CAR string, DIDs, OIDC claims, domains/levels/tiers      │
└──────────────────────────────────────────────────────────┘

Layer 1: Identity & Trust Primitives

Purpose: Establish WHO the agent is and WHAT it claims to do.

Component	Standard	Purpose
CAR String	CAR Spec	Compact agent identifier
DID	W3C DID Core	Cryptographic identity
OIDC Claims	OpenID Connect	Token-based auth
Domains	CAR Spec	Capability areas (10 codes)
Levels	CAR Spec	Autonomy degree (L0–L7)
Tiers	CAR Spec	Trust classification (T0–T7)

Layer 2: Capability Certification & Extensions

Purpose: Verify and certify WHAT the agent can actually do.

Component	Purpose
Attestations	Third-party capability proofs
Certification Authority	Issues/revokes attestations
Trust Engine	Computes runtime trust scores
Extensions	Optional governance hooks
Registry API	Agent discovery and management

Layer 3: Semantic Governance & Runtime Assurance

Purpose: Monitor HOW the agent behaves and enforce guardrails.

Component	Purpose
Instruction Integrity	Bind agents to approved instruction sets
Output Schema Binding	Prevent data exfiltration
Behavioral Drift Detection	Detect anomalous behavior
Context Authentication	Prevent indirect injection
Audit Trail	Immutable action history

System Components

Agent Naming Service (ANS)

The registry for agent registration, discovery, and DID resolution.

POST /api/v1/agents          — Register
GET  /api/v1/agents?domain=F — Discover
GET  /api/v1/agents/{did}    — Resolve

Certification Authority (CA)

Issues verifiable attestations that certify an agent's capabilities.

Trust Engine

Continuously evaluates trust scores from three inputs:

Trust Score = Certification(30%) + Behavior(40%) + Context(30%)

Cognigate Policy Engine

Evaluates governance policies before agent actions:

Agent Action → Cognigate Policy Check → Allow/Deny → Execute/Block

Agent Lifecycle

1. Registration  → CAR string assigned, DID created
2. Certification → Capabilities attested by CA
3. Deployment    → Runtime tier computed for context
4. Operation     → Actions gated by trust + policy
5. Monitoring    → Behavior tracked, trust updated
6. Re-Cert       → Periodic re-attestation
7. Revocation    → Immediate trust zeroing if needed

Agent Provenance

How an agent was created affects its initial trust:

Creation Type	Trust Modifier	Use Case
FRESH	±0	New agent, starts at baseline
CLONED	-50	Copy of existing agent
EVOLVED	+100	Upgraded from proven agent
PROMOTED	+150	Elevated from lower tier
IMPORTED	-100	External agent, untrusted source