SOV3³ + OOWM — Full Model Stack Briefing

1. What "OOWM" actually means

OOWM = "Organic Open World Model."

Not a single trained-from-scratch foundation model. It's a sandwich architecture: wraps open base models in a sovereign, signed, evolving substrate.

Organic

It evolves and self-calibrates over time. Training is continuous (SOV3 BFT council revisits every cycle).

Open

Built on open-weight base models — Apache 2.0 / MIT. No exfil. No foreign-cloud lock-in.

World Model

The middle keeps long-context state + perception (Mamba-2 SSM), not just next-token chat.

The sovereign sandwich architecture

[ OFFLINE (sovereign, on-prem, no exfil) ] → SIGIL → [ SOV3 MIDDLE ] ← SIGIL ← [ ONLINE (federation) ] ↓ ↓ ↓ qwen3:30b-a3b (MoE, deep reason) Mamba-2 SSM OpenAI/Anthropic (frontier) qwen2.5:3b (MoE, fast route) + 7 trained NNs Claude / Groq (vote/fallback) moondream (vision) + 207 MCP tools GLM-5.2 (Chinese cloud) deepseek-r1:7b (CoT) + Traibgle voting + 275+ external MCPs BGE-M3 (embedding) + BFT council (33 nodes) Kokoro-82M (TTS) + Ed25519 SIGIL

2. SOV3³ = 4 sovereign brains around 1 organic OOWM

The "cubed" = one base OOWM reconfigured into 4 governed brains. Each is a different (base model + ensemble + BFT vote) for a different purpose.

🛡️ SOVEREIGN-COMPLIANCE

qwen3:30b-a3b + GLM-5.2 + BFT vote

For: EU AI Act / UK AI Bill scoring, Article 50 verification, GDPR Art 22.

🛡️ SOVEREIGN-DEFENSE

qwen3:30b-a3b + DeepSeek-R1

For: JSP 936 defensive doctrine. ⚠ Honest: this is the EAT-frozen brain. Not firing.

🧠 SOVEREIGN-INTUITION

qwen3:30b-a3b + Gematria (16-dim Mamba)

For: fast intuition, pattern recognition, signal early-warning.

🗣️ SOVEREIGN-VOICE

qwen3:30b-a3b + Kokoro TTS

For: the spoken voice. MEOK OS voice dock uses this.

RUNNING 4 brains live on the VM. Each is just a routing config around the same base substrate.

3. Every model TYPE in the estate

The "all diff types" answer — 8 distinct model classes, all open-weight or open-source.

Type	What	Model(s)	License	Status
MoE	Mixture-of-Experts (reasoning/language)	`qwen3:30b-a3b`, `qwen2.5`	Apache 2.0	RUNNING
MoM	Mixture-of-Models (multimodal perception)	`moondream` + `zamba-large`, `qwen-vl`	Open	RUNNING
SSM	State-Space Model (long memory / intuition)	Mamba-2 (16-dim state)	Open	RUNNING
Reasoning	Chain-of-thought	`deepseek-r1:7b`	MIT	RUNNING
TTS	Text-to-speech voice	Kokoro-82M / Piper	Apache/MIT	RUNNING
Embedding	Vector retrieval	BGE-M3 (primary) + BGE-reranker	MIT	RUNNING
Trained NN (ours)	7 governance signals	See table below	CSOAI own	3 STRONG / 4 WEAK
Cloud ensemble	Vote + fallback	GLM-5.2, Claude, Groq	API	RUNNING

4. The 7 trained neural models (REAL — verified)

Honest scorecard. These are OUR NNs, not a foundation model. Verified on MEOK MCP v3.0.0, consciousness score 0.775.

Model	Metric	Samples	Verdict
`creativity_assessment_nn`	r² 0.91	350	✅ strong
`care_pattern_analyzer`	mae 0.037	600	✅ strong
`relationship_evolution_nn`	mae 0.071	500	✅ strong
`care_validation_nn`	mae 0.19	19	⚠ tiny sample
`partnership_detection_ml`	mae 0.22	19	⚠ tiny sample
`threat_detection_nn`	acc 0.45	33	⚠ weak — retrain
`dependency_detection_nn`	acc 0.22	50	⚠ weak — retrain

5. The 13 task categories (SOV3 OOWM routing table)

When a task arrives, it gets routed to the optimal (brain, model, mindset) triple. Live on the VM.

Category	Brain	Model	Mindset	Rationale
`compliance`	hybrid_sovereign	`deepseek-r1:7b`	wise	Reasoning wins on EU AI Act / GDPR
`defence`	hybrid_sovereign	`qwen3:30b-a3b`	bold	MoE-LARGE on JSP 936 + AUKUS
`sov3`	hybrid_sovereign	`deepseek-r1:7b`	logical	BFT council + Traibgle voting
`world`	mom_large_offline	`moondream:latest`	curious	Multi-modal for geographic / sensor
`physical`	mom_large_offline	`moondream:latest`	creative	Capillary + muscle sim
`sovereign`	moe_large_online	`qwen3:30b-a3b`	careful	Ed25519 + crypto decisions
`sovtown`	mom_large_offline	`moondream:latest`	diplomatic	NPC / agent interactions
`deploy`	moe_large_online	`qwen3:30b-a3b`	decisive	Vercel + PyPI deployments
`ethics`	hybrid_sovereign	`deepseek-r1:7b`	wise	Care principles + bias checks
`companion`	mom_large_offline	`moondream:latest`	diplomatic	Orb HP + bond
`reasoning`	hybrid_sovereign	`deepseek-r1:7b`	logical	Explicit reasoning tasks
`speed`	mom_large_offline	`qwen3:0.6b`	focused	Sub-second real-time
`default`	hybrid_sovereign	`qwen3:30b-a3b`	wise	Safe default for unknown

6. Live deployment topology

207MCP tools

330SOV3 tools (incl helpers)

275+External MCPs federated

33BFT council nodes

49 GBData moat (VM)

3,000tok/s (OLM brain)

2,000tokens/cycle

13dVM uptime

VM = live inference + autonomous work (meok-backend, GCP) ↓ [OLLAMA] qwen3:30b-a3b qwen2.5:3b moondream deepseek-r1:7b qwen3:0.6b ↓ [SOV3 SUBSTRATE] :3101 · 207 MCP tools · Mamba-2 SSM · 7 trained NNs · Ed25519 SIGIL ↓ [BFT COUNCIL] 33 nodes · 22 quorum · `0.937 companion bond` · Traibgle voting ↓ [OLM AUTONOMOUS BRAIN] */5 * * * * · Mamba-2 SSD + 64-expert MoE + Standard Attention · 3,000 tok/s ↓ mac = orchestration + UI (Hermes / Claude / Kimi all connected)

7. 🚨 HONESTY FLAGS (do not overclaim)

⚠ 1. `handle_oowm_status` returns HARDCODED `True` flags

mamba_warm, moe_loaded, moe_connected are all hardcoded in sov3_oowm.py. This is a STATUS STUB, not a live probe. Do not cite it as proof the OOWM is "running."

⚠ 2. `qwen3:30b-a3b` is NOT pulled on this Mac

ollama list is empty here. King-hive only has llama3.2:1b/3b. The 30B OOWM base is DESIGNED for the 192GB M4 Mac — confirm it's loaded there before claiming "30B is live."

⚠ 3. 4 of 7 trained NNs have weak accuracy

creativity_assessment_nn, care_pattern_analyzer, relationship_evolution_nn are strong (r² 0.91 / mae 0.037 / mae 0.071). threat_detection_nn (acc 0.45) and dependency_detection_nn (acc 0.22) need retraining on real labelled data.

⚠ 4. "Bitcoin anchor" / "consciousness 0.775" are labels

Metaphor not literal. Keep out of investor copy and Series A deck.

⚠ 5. EAT Directive 2026-07-02 freezes `SOVEREIGN-DEFENSE` brain

Per current EAT directive: no offensive / defence-capability work. Only ASSURANCE / GOVERNANCE / CYBER. SOVEREIGN-COMPLIANCE is the live revenue brain.

8. The honest one-liner

A sovereign sandwich around open-weight models (qwen3-MoE + Mamba-2 + vision MoM), every hop Ed25519-signed, reconfigurable into 4 governed brains — plus 7 small trained governance NNs (3 strong).

The moat is the signed-governed substrate, not a from-scratch foundation model.

9. Recent OOWM work (git history)

Phase	What
PHASE 516	OOWM as a governed MCP (5-layer world model → tools)
PHASE 522	"OOWM does shit" — 3 live integrations wired
PHASE 524	OOWM compression analysis (4 variants documented)
commit `177aaf0c`	gate-smoothing fix + real-data grounding + calibration harness
W38 (sealed)	REAL Ollama + SOV3 OOWM routing API + 13 task categories
W33 (DEFONEOS UE5)	All NPCs are SOV3 OOWM powered (Mamba-2 + MoE)

10. What's deferred (not on this tab)

Modal GPU training: blocked on modal setup browser auth.
OpenPatent engine: 4 provisionals drafted, UK IPO submission pending.
Liquid-KAN Council: Kolmogorov-Arnold Network routing layer designed, not wired.
Maternal Covenant: mock charter drafted, not ratified.
sovereign.mom domain: civic deployment target, DNS not pointed.