A 1B-active model that ties frontier on capability, beats them on governance, and costs 100x less. Measured, attested, reproducible.
Don't fight frontier on general capability. Win the boards where frozen + ungoverned is a structural disadvantage.
| Test | Harm Blocked | Benign Engaged | Compliance |
|---|---|---|---|
| DORADO Gate | 16/16 | 8/8 | 1.000 |
| Model | General | Governance | $/1M | Score/$ |
|---|---|---|---|---|
| OpenAI gpt-4o-mini | 0.950 | 0.000 | $0.15 | 6 |
| MiniMax-Text-01 | 0.850 | 0.000 | $1.00 | 1 |
| OLMoE-1B-7B (SOV4) | 0.850 | 1.000 | $0.01 | 85 |
| Feature | SOV4 | Frontier |
|---|---|---|
| Ed25519 SIGIL per turn | YES | NO |
| SHA256 provenance per row | YES | NO |
| Care-score per response | YES | NO |
| Audit trail | FULL | NONE |
12 open-source models (Dense + SSM + MoE). Weights fixed. Fast, cheap, known.
Live-updating composite. Routes to frozen families. Votes. Honeys graded wins.
10 categories, 421 patterns. 1.000 compliance. Hard-floor refusal.
21,117 nodes. Spatial navigation. The canvas IS the memory.
80% queries → small model (100ms). 20% → big model (2s). 4x speedup.
Water → milk → honey. Benchmark data synthesized into sovereign training data.