AI Governance Architecture

TealTiger is a runtime governance framework for autonomous AI agents. It enforces policy at execution boundaries, produces cryptographic evidence, and maintains separation of duties between governance teams and development teams.

---

​

Architecture Overview

---

​

Three-Stage Evaluation Pipeline

Every request flows through three stages:

​

1. Pre-Evaluation (Sequential, Short-Circuit)

Evaluated in order. If any check fails, the pipeline stops immediately with DENY.

​

2. Module Evaluation (Parallel)

All registered modules evaluate concurrently via Promise.allSettled. Results are merged using most restrictive wins — the highest-severity action becomes the final decision. 10 governance domains, each with independent modules:

​

3. Post-Evaluation (Async, Non-Blocking)

After the decision is computed, evidence is emitted without blocking the response:

• TealProof — Append decision hash to Merkle tree
• SIEM Export — Emit structured event (JSON/CEF/LEEF)
• Response Hooks — Invoke webhooks for violations
• OTel Spans — Emit governance telemetry

---

​

Separation of Duties

The architecture enforces a clear boundary between who defines governance and who implements applications: This is enforced technically — not just organizationally. Forbidden config keys are rejected. FREEZE rules are immutable. Bundle signatures are verified on every load.

---

​

Evidence Layer

Every decision produces a TEEC v2.0.0 evidence envelope containing:

• Decision action + reason codes + risk score
• NHI identity (who triggered the action)
• Cryptographic proof (Merkle root + inclusion proof)
• Policy version + bundle hash
• Automation level + control ID + OWASP category
• Cost evidence (tokens, estimated cost)
• Correlation ID (OpenTelemetry-compatible)

Evidence is exported to SIEM, stored as Merkle receipts, and aggregated into Governance Passports for continuous coverage proof.

---

​

Design Principles

---

​

Related Documentation