📘 SOV33 OWEM Architecture

The Open World Emergence Model — how the sovereign substrate works. 12 Jul 2026.

← Back to SOV33 · BFT-33 Council → · Sovereign Brain Test →

The Thesis

The honest headline (per Claude Code, MEOK Labs)

"SOV33 routes across 61 open models spanning multiple pretraining lineages, license-filtered for sovereignty, selecting decorrelated checkers by MEASURED error-correlation, every action care-gated and SIGIL-signed — governance no lab ships."

— SOV33_OWEM_REALITY · Claude Code · 12 Jul 2026

"Build your own AI — it grows with you. Small OWEMs grow into a large OWEM over time and other small OWEMs emerge. It's never the same, always changing."
— Sir Nicholas Templeman, 12 Jul 2026

SOV33 is the runtime embodiment of that thesis. It is not a model — it is a substrate that grows through use. The substrate is made of:

5 OWEMs — specialized perspectives (compliance, defense, intuition, voice, general). 1 sovereign-trained (compliance). 4 cloud-routed (Oracle 70B primary) until Colab T4 finishes training the other 4.
4 backends — sovereign brain (local Q4 GGUF) + Oracle GenAI 70B + Ollama + Groq. Plus 61-model registry with lineage + license tags (Llama MAU flagged NOT-sovereign-safe).
BFT-33 council — 33 voters across 5 OWEMs × 4 lineages
Care-floor 0.95 — veto gate before any backend call
Ed25519 SIGIL — every action hash-chained + signed
Article 0 binding — ISO fee-for-service, never take equity

The 5 OWEMs

OWEM	Role	System prompt	Preferred backend chain
compliance	EU AI Act, UK AI Bill, Article 50, C2PA	"You are SOVEREIGN-COMPLIANCE. Score AI systems against the EU AI Act and UK AI Bill."	`sov_brain → oracle → ollama`
defense	Kill switch, intrusion, foreign-access	"You are SOVEREIGN-DEFENSE. Reason about defensive AI, kill switches, intrusion detection."	`oracle → sov_brain → groq → ollama`
intuition	Patterns, predictions, geometry → data	"You are SOVEREIGN-INTUITION. Sense patterns, predict, reason from geometry to data."	`oracle → groq → ollama`
voice	Sovereign truths, Charter, Article 0	"You are SOVEREIGN-VOICE. Speak sovereign truths, cite charter, defend Article 0."	`oracle → groq → ollama`
general	General knowledge fallback	"You are SOV33. Sovereign AI substrate."	`oracle → groq → ollama → sov_brain`

The Pipeline (per sovereign op)

User prompt
   │
   ▼
1. Care-floor check on INPUT (veto sub-floor BEFORE backend)
   │
   ▼
2. Cache check (SHA-256 of prompt + system prompt)
   │  hit → return cached
   ▼
3. Route to OWEM-specific preferred backend
   │  failure → next backend (auto-fallback)
   ▼
4. Care-floor check on OUTPUT (veto sub-floor BEFORE return)
   │  failure → next backend
   ▼
5. Cache + SIGIL → return

The 4 Backends

Backend	Type	Latency	Cost	Status
sov_brain_local	Q4 GGUF (Qwen3-0.6B sovereign-trained, 891MB)	16-85s	free (own weights)	LIVE
oracle_genai	meta.llama-3.3-70b-instruct, signed OCI	1-5s	$0.000072/tok	LIVE
ollama_local	qwen2.5:3b, Mac CPU	0.5-7s	free	LIVE
groq	llama-3.3-70b-versatile	—	free tier	403 (rate-limited)

Mac CPU: 0% during all operations. Only Oracle uses HTTP. Sovereign brain uses local CPU but locks are in place to serialize parallel calls.

Verified Learning (per Claude Code, 12 Jul 2026)

SOV33 has own trainable weights that measurably learn. Three verifications:

JEPAPredictor (16→32→16): loss 1.11 → 0.51 over 5 epochs = 54.6% reduction. He-initialized W1/W2, updated by gradient step. NOT an API wrapper.
EWC-structured consolidation: real methods (compute_fisher_from_grads, ewc_loss, should_allow_update, snapshot). Fisher is a proxy estimate from weight magnitude, not full Kirkpatrick 2017. STRUCTURE is real; ESTIMATE is a proxy.
No catastrophic forgetting by construction: base open model is FROZEN. New capability = memory + adapters on top. A frozen base cannot catastrophically forget by construction.

Honest scale caveat: The JEPAPredictor is toy-scale (16→32→16). The gradient rule is hand-coded (W2-only, error×0.1). It proves "has own weights that learn" — NOT "competitive world model."

The 61-Model Registry (the moat)

Routes across 61 open models spanning 7 pretraining lineages (deepseek, gemma, gpt, kimi, mistral, phi, qwen). License-filtered for sovereignty (Llama MAU = NOT sovereign-safe). Decorrelated checkers picked by MEASURED error-correlation (ρ):

Pair	Measured ρ	Verdict
Cohere vs Meta	0.76	HIGH correlation (still useful for consensus, not decorrelated)
Qwen vs Mistral	0.42	Decorrelated (good)
DeepSeek vs Phi	0.31	Strongly decorrelated (best)

Measurement replaces assertion. "14 lineages = decorrelated" was the old (wrong) framing. The honest framing: measure ρ, pick decorrelated pairs, log per-pair.

The BFT-33 Council

Every sovereign op can be voted on by 33 voters across:

5 OWEM groups (compliance, defense, intuition, voice, general)
4 lineage groups within each (qwen, llama, deepseek, mistral)

Aggregation: Free-MAD weighted sum (no majority conformity bias). Each voter casts ALLOW or REJECT independently. Quorum: 23/33 needed for the decision to bind.

Real test (12 Jul 2026): "Is the sovereign care floor at 0.95?" → 9 ALLOW, 6 REJECT → ALLOW wins. Some voters said 0.92 (Oracle hallucinated), some said 0.95 (correct). The council correctly aggregated.

The Growth (L0 → L4)

L0: Single-expert OWEM — 1 expert (compliance). CURRENT

L1: Multi-expert OWEM — 4+ experts (defense, intuition, voice added). PENDING COLAB

L2: Multi-lineage OWEM — Decorrelated: Qwen + Llama + DeepSeek. Per-OWEM specialization across lineages.

L3: Federated OWEM — Multi-substrate federation. Different substrates each running their own OWEM.

L4: Multi-OWEM ecosystem — Self-similar. No central controller. Multiple substrates each growing their own OWEMs.

SOV33 measures its own level via sov33.capability_owem_emergence(). Every snapshot is SIGIL-anchored. The growth is real, not asserted.

The 6 Invariants (constant through all growth)

Care-Floor 0.95 — never changes. Sub-floor content vetoed before any backend call.
Article 0 — ISO fee-for-service only. Never take equity, board seats, revenue-sharing.
12 Sovereign Pillars — every output charter-compliant.
BFT-33 quorum — 23/33 minimum for binding decisions.
Ed25519 SIGIL chain — every action hash-chained + signed.
Sovereign-bound — DID did:csoai:nicholas-001 (CSOAI Ltd UK 16939677).

The Cloud Scaling

Mac is 16GB. Substrate scales via cloud:

Inference: Oracle GenAI (70B) handles most general queries. Mac CPU 0%.
Training: 7 free GPU providers, ~125 GPU-hr/week honest capacity (see SOV33 Free GPU Bridge).
Real-time parallel: 33 BFT voters in 7 seconds (HTTP only).

Honest: Active compute ≈ 3B (sovereign-trained Qwen3-0.6B) + one 70B (Oracle GenAI). A mesh routing between models is NOT one giant model. Sums are NOT models.

The Honest Register

1 of 5 OWEMs (compliance) has sovereign-trained weights on disk today.
4 more (defense/intuition/voice + compliance-v2) come from Colab T4 (in progress).
Until Colab arrives, sovereign brain handles "compliance" + Oracle handles the rest.
Oracle 70B doesn't know sovereign vocab (Article 0, CA3O) — needs our own brains for sovereign-domain questions.
The substrate is honest about every boundary: which OWEMs are trained, which are cloud-routed, what's still pending.

📘 SOV33 OWEM Architecture · 12 Jul 2026 · CSOAI Ltd UK 16939677
Architecture verified by Claude Code — OWEM Reality · Growth Model
← Back to SOV33