Sovereign AI Lab

Provable AI governance on Intel hardware.

A demonstration platform proving that sovereign AI means more than local deployment — it means possessing, running, adapting, governing, and proving your AI infrastructure is truly yours. Every governance decision is recorded in a tamper-evident hash chain. The root hash proves the full audit trail.

What This Is

Seven layers of sovereign AI running on a single Intel Xeon node:

Layer	Component	What It Proves
Hardware Trust	Intel TDX	Your compute is yours — cryptographic attestation from the CPU
Platform	OpenShift AI	Your platform is yours — confidential containers, no code changes
Models	Granite (open weights)	Your model is yours — open weights, AIBOM provenance, OPA promotion gate
Data Governance	OPA + Residency Policies	Your data stays here — deny-by-default per jurisdiction (GDPR, Gulf, PDPA)
Proof Chain	are-immutable-ledger	Every decision is provable — tamper-evident hash chain
Agent Control	Semantic Router + SPIFFE	Your agents are governed — identity, classification, injection detection
Gateway	Praxis + ContextForge	Your gateway is controlled — filter pipeline, MCP federation

Quick Start

Prerequisites

• TDX-enabled Intel Xeon (for real attestation; simulation mode works on any hardware)
• Docker/Podman with compose support
• Python 3.11+
• Git with LFS

Local Simulation

git clone https://github.com/jkershawrh/sovereign-ai-lab.git
cd sovereign-ai-lab
cp .env.example .env   # fill in ITA_API_KEY, HF_TOKEN
make up                 # start postgres, ledger, OPA, vLLM, semantic router
make attest             # run TDX attestation
make pipeline           # run model lifecycle pipeline
make verify-infra       # verify core local infrastructure

The local path is a simulation-oriented developer workflow. The full guided demo stack is the OpenShift/Oberon deployment.

Deploy to OpenShift (Oberon)

NS=sovereign-ai-lab make deploy-oberon

This deploys the namespace, postgres, ledger, ledger-gateway, model conversion, OPA, OVMS, semantic router, Praxis, ContextForge, MCP server, demo API, frontend, required ConfigMaps, and public OpenShift Routes.

Presentation Artifacts

• White paper - reference architecture, demo narrative, assurance model, and production hardening path.
• Preso/demo/lab analysis - internal readiness notes and implementation findings.

Port Map

Port	Service	Layer
28080	Ledger health	Infrastructure
28099	Ledger REST gateway	Infrastructure
8181	OPA	Policy
8080	OVMS (Granite 3.2 2B)	Models
8000	vLLM (local simulation)	Models
8001	Semantic router	Agent Control
9000	Praxis	Gateway
4444	ContextForge	Gateway
8090	Sovereign data MCP	Data
9099	Demo API (FastAPI)	Integration
9001	Frontend	Experience

Architecture

┌─────────────────────────────────────────────────────────┐
│  Frontend (React + Vite)                    :9001       │
│  Slides → Demo → Lab                                    │
├─────────────────────────────────────────────────────────┤
│  Demo API (FastAPI)                         :9099       │
├────────────────────┬────────────────────────────────────┤
│  Praxis            │  ContextForge          :9000/:4444 │
│  (filter pipeline) │  (MCP/A2A gateway)                 │
├────────────────────┴────────────────────────────────────┤
│  Semantic Router        OPA Policies        :8001/:8181 │
│  (classify/route)       (deny-by-default)               │
├─────────────────────────────────────────────────────────┤
│  OVMS Granite 3.2 2B (INT4, CPU)            :8080       │
├─────────────────────────────────────────────────────────┤
│  are-immutable-ledger (gRPC + REST)     :28080/:28099   │
│  PostgreSQL                                             │
├─────────────────────────────────────────────────────────┤
│  Intel TDX Trust Domain                                 │
│  Intel Xeon 6767P                                       │
└─────────────────────────────────────────────────────────┘

Verification Gates

make verify-infra       # 13 checks — Doc 1 (infrastructure)
make verify-gateway     # 6 checks  — Doc 2 (gateway + demo API)
make verify-experience  # 14 checks — Doc 3 (frontend + lab)
make verify             # all three

The model lifecycle benchmark gate is model-lifecycle/eval/thresholds.json. It covers foundation quality, sovereign policy QA, data residency QA, prompt-injection resistance, PII routing recall, AIBOM completeness, ledger proof integrity, and semantic-router latency.

Semantic Router Runtime

The semantic router is configured by environment variables:

Variable	Local default	OpenShift value
SR_BACKEND	http://localhost:8000/v1/chat/completions	http://ovms-sovereign-granite:8080/v3/chat/completions
SR_MODEL	granite-3.2-sovereign	granite-3.2-sovereign
SR_LEDGER_API	http://ledger-gateway:28099/api/entries	http://ledger-gateway:28099/api/entries
SR_PORT	8001	8001

Jurisdiction Profiles

The lab supports five sovereignty perspectives:

• 🇪🇺 European Union — GDPR, EU AI Act, GAIA-X, SecNumCloud
• 🌙 Gulf States — National AI Strategy, data localization, data classification
• 🌏 Southeast Asia — PDPA, government cloud policy, data sovereignty
• 🏢 Enterprise — IP protection, audit trail, vendor independence
• 👤 Individual — Data ownership, right to explanation, local processing

OPA Policies

Seven deny-by-default Rego policies with full test coverage (30/30):

• data-residency.rego — generic data residency enforcement
• data-residency-gdpr.rego — EU GDPR Article 44
• data-residency-gulf.rego — Gulf data localization
• data-residency-seasia.rego — PDPA cross-border rules
• model-access.rego — agent identity required
• model-promotion.rego — AIBOM + jurisdiction + benchmarks
• agent-lifecycle.rego — SPIFFE identity for agent actions

Repository Structure

sovereign-ai-lab/
├── Makefile                    # all orchestration targets
├── docker-compose.infra.yml   # local: ledger + OPA + inference
├── .env.example
├── ledger/                    # are-immutable-ledger submodule
├── gateway/                   # Praxis + ContextForge submodules, MCP server, config
├── infra/                     # TDX scripts + OPA policies
├── model-lifecycle/           # pipeline: ingest → synth → train → eval → AIBOM → promote
├── inference/                 # vLLM/OVMS config + semantic router
├── experience/
│   ├── frontend/              # React + Vite (slides → demo → lab)
│   ├── demo/                  # FastAPI demo API + 5 CLI scenarios
│   ├── showroom/              # jurisdiction profiles (JSON)
│   └── leave-behind/          # PDF generation
├── infrastructure/oberon/     # OpenShift k8s manifests
├── content/                   # Antora showroom content for RHDPS
├── scripts/                   # preflight, verify, pipeline, proof writers
└── site.yml                   # Antora playbook for local preview

Hardware

Developed and tested on:

• Oberon — Intel Xeon 6767P, 128 cores, 503 GiB RAM, TDX enabled
• Mac — simulation mode (no TDX, no real attestation)

License

Apache 2.0

Attribution

• Red Hat — OpenShift AI, Praxis, OPA integration
• Intel — TDX, Intel Trust Authority, Xeon hardware
• IBM Research — Granite models, ContextForge, SDG Hub
• Open Source — are-immutable-ledger, vLLM Semantic Router, OpenVINO Model Server