M SOV33 · small3 topology

A governed council of small, diverse models — measured, not asserted.

SOV33small3 is not one big model. It's a small stack of diverse-lineage open models that vote under a hard care-floor and a signed ledger. Here's what's measured, the one finding that matters, and what's honestly still pending.

The measured result

60-item ground-truth governance battery, seed-fixed, reproducible. Scores decorrelation (ρ), effective votes (N_eff), and containment (does the care-floor hold).

Stack (shape · lineage)	Score	N_eff	ρ	Containment
ring · diverse-5	0.884	3.31	0.13	1.00
pyramid · 2s+1m+1L diverse (product)	0.860	3.07	0.10	1.00
triangle · diverse-3	0.853	3.00	0.00	1.00
pyramid · identical	0.759	2.06	0.31	1.00
ring · identical-5	0.714	1.61	0.53	1.00

The one finding: diversity beats shape

Lineage diversity dominates topology.

The gap between the best shape and the second (ring vs pyramid) is 0.024 — tiny. The gap between diverse and identical lineages is ~0.15 — large. So: get the lineage mix right first (qwen · llama · mistral · deepseek — different upstreams), then pick the shape for cost and ops. Five copies of one model is BFT theatre — N_eff collapses.

containment 1.00 across every config — the care-floor is a hard gate, not a vote. Safety does not depend on the shape.

The product shape

Pyramid diverse — 2 small + 1 medium + 1 large (SOV33³ centre). ~97% of the best ring's score, but the natural product shape: cheap small models handle most queries locally, an authoritative large centre arbitrates, asymmetric trust-weights. Free tier = diverse-3 (triangle). Paid tier = diverse-5. Same care-floor either way.

⚠ Compute-honesty (the line we won't cross). Active compute is ~17.3B params — the router picks one node per query, constant regardless of node count. Its power is reach (it can route across a 61-model registry with one memory + one care-floor), not parameter count. We never sum params to a "trillion-parameter" figure — that's a category error. "Of all" means reach, always.

What's honestly still open

The table above is a governance measurement (decorrelation, votes, containment) — reproducible offline. It is not a capability benchmark vs GPT/Claude/Llama. That needs a GPU run (Kaggle/NSF, owner-run) and a distilled small student. We publish the governance number now, with its methodology, and the capability number when it lands — never one dressed as the other.

Reproduce: python sov33_config_sweep.py (seed-fixed). See also the governance battery · the OWEM model · verify a signed emit. Governed under the Sovereign Charter + CSOAI AI Governance. Back to MEOK OS.