Case Study: Sovereign AI Offline Mode

This chapter covers APR's Sovereign AI capabilities, particularly the --offline mode that enables air-gapped deployments.

Overview

Sovereign AI refers to AI systems that are fully controlled, operated, and audited by the user, without reliance on centralized APIs or proprietary cloud infrastructure.

Per Section 9.2 of the specification:

HARD REQUIREMENT: The system must be capable of operating continuously in an "Air-Gapped" environment (no internet connection) once necessary artifacts are acquired.

Compliance Checklist

Using Offline Mode

Basic Usage

# Run a model in offline mode (production recommended)
apr run --offline model.apr --input data.json

# Offline mode rejects uncached remote sources
apr run --offline hf://org/repo  # ERROR: OFFLINE MODE

Caching Models First

# Step 1: Import model to cache (requires network)
apr import hf://TinyLlama/TinyLlama-1.1B -o tinyllama.apr

# Step 2: Run in offline mode (no network required)
apr run --offline tinyllama.apr --input prompt.txt

Model Source Types

APR supports three model source types:

Network Isolation

The inference loop is designed to be physically incapable of network IO:

1. No std::net imports in inference code
2. No reqwest or HTTP client libraries
3. No hyper or async networking
4. Type-system enforced isolation

Verification

Run the V11-V15 tests to verify network isolation:

cargo test --test spec_checklist_tests v1

Example Code

//! Sovereign AI: Offline Mode Example
//! Run: cargo run --example sovereign_offline

use std::path::PathBuf;

fn main() {
    println!("=== Sovereign AI: Offline Mode Demo ===\n");

    // Demonstrate source types
    let sources = [
        ("model.apr", "Local"),
        ("hf://org/repo", "HuggingFace"),
        ("https://example.com/model.apr", "URL"),
    ];

    for (source, source_type) in sources {
        println!("{} -> {}", source, source_type);
    }

    // Offline mode behavior
    println!("\nOffline Mode:");
    println!("✅ Local files: Always allowed");
    println!("✅ Cached models: Allowed");
    println!("❌ Uncached HF: REJECTED");
    println!("❌ Uncached URLs: REJECTED");
}

Run the example:

cargo run --example sovereign_offline

Cache Structure

Models are cached in ~/.apr/cache/:

~/.apr/cache/
├── hf/
│   ├── openai/whisper-tiny/
│   └── TinyLlama/TinyLlama-1.1B/
└── urls/
    └── <hash>/  (first 16 chars of URL hash)

Production Deployment

For production deployments:

1. Pre-cache all models during deployment
2. Always use --offline flag
3. Verify network isolation with integration tests
4. Air-gap the inference environment if required

# Deployment script
apr import hf://org/model -o /models/model.apr
chmod 444 /models/model.apr  # Read-only

# Runtime
apr run --offline /models/model.apr --input request.json

Popperian Falsification

The offline mode implementation includes Popperian falsification tests:

References

• Section 9.2: Sovereign AI Compliance
• Local-First Software (Kleppmann et al., 2019)
• Example: sovereign_offline.rs