Runtime Governance Reference Architecture

This page describes the canonical runtime governance architecture for TealTiger v1.3. It shows how governance is authored, distributed, enforced, and evidenced across an enterprise AI deployment.

Architecture Layers

Layer	Responsibility	Who Owns It
Policy Authoring	Define rules, FREEZE controls, NHI scopes, cost ceilings	Governance Team
Bundle Distribution	Sign, publish, and version policy bundles	Governance Team
Runtime Enforcement	Evaluate requests, enforce decisions, emit evidence	TealEngine (in application)
Evidence & Audit	Cryptographic receipts, SIEM export, OTel spans	Automated (post-evaluation)
Platform Integration	Native adapters for cloud agent services	Platform Team

Layer 1: Policy Authoring

Governance teams author policies in YAML/JSON, define FREEZE rules, configure NHI scopes, and set cost ceilings:

# governance-bundle/production.yaml
version: "3.2.0"
freeze_rules:
  - id: FREEZE-001
    condition: { action: "database.drop" }
    reason: FREEZE_DATABASE_DROP

rules:
  - condition: { action: "payment.*", amount: { gt: 100 } }
    automation: approval_required
  - condition: { action: "*.read" }
    automation: auto_allow

nhi:
  enforce_scopes: true
  require_attestation: true

cost_governance:
  per_request_max: 2.00
  daily_per_agent_max: 50.00

Tooling: Policy authoring CLI, golden corpus testing, schema validation.

Layer 2: Bundle Distribution

Governance teams sign bundles with Ed25519 and publish to an internal registry:

Governance Team → Sign (Ed25519) → Publish to Registry → Applications Pull at Startup

Distribution Channel	Use Case
S3 / GCS / Azure Blob	Cloud-native, IAM-controlled
OCI Registry	Container-native artifacts
npm / PyPI private registry	Developer-familiar
Git repository	Version-controlled, PR-based

Integrity guarantee: Applications verify bundle signature on every load. Tampered bundles are rejected.

Layer 3: Runtime Enforcement

TealEngine v1.3 runs inside the application process. Every agent action passes through the governance pipeline before execution:

Agent Action → TealEngine.evaluate() → Decision → Application handles ALLOW/DENY/PENDING

Enforcement Points

Point	What’s Evaluated	Example
Tool invocation	Tool name, parameters, agent scope	`web_search`, `file.write`, `database.query`
Model call	Model selection, cost estimate, content	GPT-4o call with 2000 tokens
Memory write	Content classification, provenance, scope	Writing customer data to shared memory
Code change	Path, branch, diff hash	Agent writing to `/src/auth/`
Session boundary	TTL, cooldown, time-of-day	Agent running for 9 hours

Deployment Options

Option	Latency	Use Case
Embedded SDK	< 25ms	Most applications (recommended)
Platform adapter	< 30ms	AWS Bedrock, AgentCore, Azure AI
Docker sidecar	< 35ms	Language-agnostic, polyglot

Layer 4: Evidence & Audit

Every decision produces evidence through multiple channels:

Channel	Format	Destination	Purpose
TEEC Envelope	JSON	Local / S3	Structured decision record
TealProof	Merkle + RFC 3161	Local / TSA	Tamper-evident proof
SIEM Export	JSON / CEF / LEEF	Splunk / Elastic / QRadar	Security monitoring
OTel Spans	OpenTelemetry	Jaeger / Datadog / Tempo	Distributed tracing
Response Hooks	HTTP webhook	PagerDuty / Slack / custom	Real-time alerting

Non-repudiation: TealProof receipts + RFC 3161 timestamps prove governance was enforced at a specific time. Third parties can verify independently using the Verification SDK.

Layer 5: Platform Integration

Native adapters for cloud agent platforms:

Platform	Adapter	Integration Point
AWS Bedrock Agents	Lambda Layer / sidecar	Action group governance
AWS AgentCore	Plugin hooks	Lifecycle governance (create/execute/tool/terminate)
Azure AI Agent Service	Azure Function / AKS sidecar	Tool-call pipeline

Each adapter maps platform-specific events to TealEngine evaluate() calls.

Separation of Duties

The architecture enforces clear boundaries:

Boundary	Governance Team	Development Team
Policy authoring	✅ Authors and signs	❌ Cannot modify
Bundle distribution	✅ Publishes	✅ Pulls at startup
Runtime enforcement	❌ Cannot deploy code	✅ Calls evaluate()
Evidence review	✅ Audits evidence	✅ Forwards to SIEM
FREEZE rules	✅ Sets immutable controls	❌ Cannot bypass
Cost ceilings	✅ Sets governance limits	❌ Cannot raise

Anti-Tamper Controls

The architecture includes multiple tamper-resistance mechanisms:

Control	What It Prevents
Bundle signature (Ed25519)	Unauthorized policy modification
FREEZE rule immutability	Disabling critical safety controls
Forbidden config keys	Setting `bypass`, `disable_enforcement`, `allow_all`
Workload identity binding	Using bundles in unauthorized environments
Capability negotiation	Loading bundles requiring unsupported features
Tamper event logging	All bypass attempts logged as security events

Compliance Mapping

Framework	How TealTiger Maps
OWASP ASI Top 10	OWASP Policy Pack covers ASI-01 through ASI-10
NIST AI RMF	Controls mapped to GOVERN, MAP, MEASURE functions
ISO 42001	AI management system alignment (A.6, A.8, A.9, A.12)
EU AI Act	Art. 9 (risk management), Art. 12 (record-keeping), Art. 14 (human oversight)
SOC 2	CC7.2 (monitoring), CC6.1 (access control), CC8.1 (change management)

Governance at Scale

Enterprise operating model in detail

Evaluation Flow

3-stage pipeline with SVG diagram

Platform Adapters

AWS and Azure native integration

Anti-Tamper Controls

Technical tamper prevention details

Getting Started

Core Concepts

Governance Domains

Policy Authoring

Architecture

Integrations

Deployment

API Reference

Cookbook

Compliance & GRC

Guides

Reference

Versions

Playground

Dashboard

Blog

Runtime Governance Reference Architecture

Architecture Layers

Layer 1: Policy Authoring

Layer 2: Bundle Distribution

Layer 3: Runtime Enforcement

Enforcement Points

Deployment Options

Layer 4: Evidence & Audit

Layer 5: Platform Integration

Separation of Duties

Anti-Tamper Controls

Compliance Mapping

Governance at Scale

Evaluation Flow

Platform Adapters

Anti-Tamper Controls

​Architecture Layers

​Layer 1: Policy Authoring

​Layer 2: Bundle Distribution

​Layer 3: Runtime Enforcement

​Enforcement Points

​Deployment Options

​Layer 4: Evidence & Audit

​Layer 5: Platform Integration

​Separation of Duties

​Anti-Tamper Controls

​Compliance Mapping

​Related Documentation

Governance at Scale

Evaluation Flow

Platform Adapters

Anti-Tamper Controls

Architecture Layers

Layer 1: Policy Authoring

Layer 2: Bundle Distribution

Layer 3: Runtime Enforcement

Enforcement Points

Deployment Options

Layer 4: Evidence & Audit

Layer 5: Platform Integration

Separation of Duties

Anti-Tamper Controls

Compliance Mapping

Related Documentation