EU AI Act Article 50 live 2 Aug 2026 โ C2PA passport ready |
Get yours โ
SOV33 doesn't try to be one giant model. Instead it has multiple OWEMs at different scales, each owning its own lane, with a governed escalation path between them.
1. Triangle (3-around-1) โ governance topology
Three small OWEMs at the vertices + one large SOV33-cubed (general) at the center as Queen/governor.
OWEM-1
lineage=qwen3-30b
brains=Compliance,Defense,Voice
OWEM-2
lineage=llama3-70b
brains=Defense,Intuition,Voice
OWEM-3
lineage=mistral-12b
brains=Compliance,Intuition,Voice
SOV33-CUBED CENTER (General OWEM โ RunPod-trained sov33-master-v2)
- Each small OWEM decides LOCALLY when confident (offline/cheap)
- Escalates open-ended / out-of-lane work to the center
- Triangle agreement is discounted by lineage ฯ (measured)
- 3 identical lineages โ 1 effective vote (escalate)
- 3 diverse lineages โ 2.3 effective votes (commit locally)
- Every cross-node message is SIGIL-signed
Built by Claude Code: sov33_triangle_owem.py. Hardened in b00c4f89 to use MEASURED ฯ (not heuristic).
2. Cascade (10/90) โ efficiency topology
LEFT brain (small fast) handles 90% of traffic ยท RIGHT brain (large deep) called only for the hard 10%.
LEFT
small fast model
~90% of traffic
handles easy/predictable queries
RIGHT
large deep model
~10% of escalations
handles hard/open-ended queries
- Difficulty estimator: regex signals + length factor
- LEFT always runs first (conscious, fast, gates via OWEM)
- RIGHT called only when difficulty > 0.5
- Every hop is SIGIL-signed and Care-Floor gated
Built by Claude Code: sov33_cascade_router.py. Honest efficiency: aggregate big, active small, every hop signed.