Sovereign Guardian
Sovereign Guardian is a family of nano-expert AI models trained to detect threats before they reach users. Every model runs entirely on-device. No cloud. No data leaves your phone.
How It Works
Guardian uses our proprietary Axiom response format — a compact field:value output designed for reliable, parseable security decisions on small models that run entirely on-device (360M / 198MB and up).
Results
| Model | Accuracy | False Positive Rate | Size |
|---|---|---|---|
| 1.7B Axiom | 89.9% | 1.6% (5-seed mean, stable) | 922MB |
Measured across 5 random seeds on a clean held-out set — mean 1.6% false-positive rate, every seed usable (no cherry-picking). We publish our real numbers and improve on them.
Domain Specialists
Guardian includes domain-specific brains evolved through our Dojo system:
- Child Protection — Detects grooming, bullying, gaming scams
- Senior Protection — Catches grandparent scams, romance fraud, voice cloning
- Financial — Identifies rug pulls, MEV attacks, phishing dApps
- General — Cross-domain social engineering detection
Each brain evolves weekly against adversarial attack scenarios, getting stronger with every generation.