🐉 SOV3³ + OOWM — Model Types Deep-Dive

Every distinct model class in the sovereign estate. What it is. What it does. How to call it. Where it runs. What's honest. Tab 3 of 3 (also see: SOV3_OOWM_TAB.html text-first, SOV3_OOWM_VISUAL.html diagram-first).

At a glance

8distinct model classes

7trained NNs (3 strong)

207SOV3 MCP tools

13task routing categories

4sovereign brains

100%open-weight or own

01 / Mixture-of-Experts

MoE

SPARSE ACTIVATION · REASONING

Mixture-of-Experts (MoE)

Deep reasoning at inference cost. Only 3B of 30B parameters active per token (A3B = "Activated 3 Billion").

Apache 2.0

What it is

An MoE replaces a single dense neural network with a router + multiple "expert" sub-networks. The router selects 2-8 experts per token, not the full set. Result: model has 30B total params but only 3B are "live" per token → runs 8-10× faster than a dense 30B model with comparable reasoning quality.

Models in the estate

Reasoning (large)	`qwen3:30b-a3b` · 18GB · 80 tok/s · 32K ctx
Fast route (small)	`qwen2.5:3b` · 2GB · ~200 tok/s · 8K ctx

Why we use it

Apache 2.0 license — fully open, no API lock-in, no exfil
Runs on M4 Mac 192GB (the sovereign target) — no GPU cluster needed
Deep reasoning for EU AI Act / GDPR / constitutional questions
Routing matches the brain selector in sov3_4_brains_1_oowm.py

Live invocation

# Local Ollama call
curl -s -X POST http://localhost:11434/api/generate \
  -d '{
    "model": "qwen3:30b-a3b",
    "prompt": "Score this AI system against EU AI Act Art 50: chatbot for NHS triage",
    "stream": false
  }' | jq .response

# Via SOV3 routing
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "sov3_apply_routing",
    "category": "compliance",
    "query": "Score triage chatbot"
  }' → routes to deepseek-r1/qwen3 depending on task

Designed for M4 192GB · confirm ollama list on the 192GB Mac before claiming live

02 / Mixture-of-Models

MoM

MULTIMODAL ENSEMBLE · PERCEPTION

Mixture-of-Models (MoM)

Multiple specialist models, one task. Vision + spatial + audio routed by input type, not by token.

Open

What it is

Where MoE uses many sub-networks inside one model, MoM uses many independent specialist whole models. Moondream sees objects in images. Qwen-VL understands spatial layouts. Zamba parses long video. A router decides which (or which combination) is needed per task.

Models in the estate

Vision (small)	`moondream` · 1.7GB · fast object recognition
Vision+spatial (large)	`zamba-large` · ~7GB · image + video captioning
Vision-Language	`qwen-vl` · understands layout, OCR, charts

Why we use it

Sovereign multimodal — no call to GPT-4V / Claude vision
Different model per modality → specialist accuracy, no compromise
Powers the "Right Brain / MOM" of the sovereign sandwich
Used in DEFONEOS UE5 NPC cognition (`UOOOWMComponent::Think`)

Live invocation

# Image analysis via MoM router
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "mom_perceive",
    "input": "image_url_or_path",
    "modality_hint": "spatial",
    "return_format": "structured_captions"
  }'
→ returns: { objects: [...], spatial_layout: ..., signed_sigil: "..." }

# Direct moondream call
curl -s http://localhost:11434/api/generate \
  -d '{"model": "moondream", "prompt": "What is in this image?"}'

Live on VM

03 / State-Space Model

SSM

LONG-CONTEXT MEMORY · INTUITION

State-Space Model (SSM)

Mamba-2 with 16-dim state. Holds the world model together — replaces the 32K context window with O(1) memory.

Open

What it is

Where transformers scale quadratically with context length (32K tokens → 1B FLOPs per token), state-space models scale linearly. Mamba-2 with state-dim 16 stores a fixed-size summary of all prior context in 16 floats — every new token updates that summary in O(1). You can run Mamba-2 over 1M tokens for the same cost as 1K tokens in a transformer.

Models in the estate

Mamba-2 (long mem)	`Mamba-2 SSM` · state-dim 16 · O(1) per token
Gematria layer	`Gematria` · 16-dim Mamba variant for intuition routing

Why we use it

The "world model" in OOWM is fundamentally an SSM — long memory, fast access
Replaces old context truncation (we were losing state every 32K tokens)
Powers SOVEREIGN-INTUITION brain (Gematria = 16-dim Mamba for fast pattern)
OLM Autonomous Brain uses Mamba-2 SSD as right-brain history layer

Live invocation

# Inside the sovereign sandwich, not user-callable directly
# Used by the OOWM engine to maintain world model state

from sovereign_temple.sov3_oowm import OOWM
oowm = OOWM(brain="sovereign-intuition")
oowm.update_state("user_just_clicked_compliance_button")
state = oowm.read_state()  # 16-dim vector, instant access, no context loss

Live (16-dim state in OOWM) · handle_oowm_status hardcodes True

04 / Reasoning LLM (Chain-of-Thought)

CoT

CHAIN-OF-THOUGHT · LOGIC

Reasoning LLM (CoT)

DeepSeek-R1 with explicit step-by-step reasoning. Wins on EU AI Act, GDPR, ethics, math, logic.

MIT

What it is

DeepSeek-R1 is trained to "think out loud" before answering. Instead of one-shot generation, it writes a thinking trace (3-15 seconds) then produces the final answer. This gives dramatically better results on tasks requiring multi-step logic — exactly the EU AI Act / GDPR / BFT council use cases.

Models in the estate

Reasoning deepseek-r1:7b · 4.7GB · 40 tok/s · 16K ctx

Why we use it

Routed to for compliance, sov3, ethics, reasoning task categories
Generates the Art 50 / GDPR gap analysis text
Writes BFT council votes with explicit reasoning trail (auditable)
Smaller (7B) → fast, fits in 16GB

Live invocation

# Reasoning call (longer wait, better answer)
curl -s http://localhost:11434/api/generate \
  -d '{
    "model": "deepseek-r1:7b",
    "prompt": "Does this AI system violate Art 14 (human oversight)? Steps: 1) read system description 2) check 4 human-oversight requirements 3) emit verdict"
  }'
→ returns: { thinking: "...step-by-step...", response: "...verdict..." }

Live on VM

05 / Text-to-Speech

TTS

VOICE · SPEECH SYNTHESIS

Text-to-Speech (TTS)

Kokoro-82M, Apache 2.0. The MEOK OS voice dock speaks through this.

Apache / MIT

What it is

Kokoro is an 82-million-parameter TTS model from hexgrad. Real-time speech on CPU. Voices: Bella, Adam, Antoni, Sarah (UK), Emma (UK), etc. We pair it with Piper as fallback. Sovereign voice — no call to ElevenLabs or OpenAI TTS.

Models in the estate

Primary	`Kokoro-82M` · 82M params · CPU real-time · 9 voices
Fallback	`Piper` · offline, monophone voices

Why we use it

Powers "Speak to your Sovereign" button on os.meok.ai
Routes through SOVEREIGN-VOICE brain (brain 4 of 4)
The 28 Jun 2026 voice-loop bug was fixed via 3-layer talk-loop resolution
Used by Voice Bridge for live verbal compliance reports

Live invocation

# Synthesis (saves wav to /tmp)
python3 -c "
from kokoro import synthesize
synthesize(
    text='This system complies with EU AI Act Article 50.',
    voice='bm_george',   # British male
    out_path='/tmp/compliance.wav'
)
"

Live (os.meok.ai voice dock)

06 / Embedding Model

VEC

VECTOR RETRIEVAL · SEMANTIC SEARCH

Embedding Model

BGE-M3 for retrieval + BGE-reranker for re-ranking. Powers every search across 49GB moat.

MIT

What it is

Embedding models convert text (or images) into fixed-dimension vectors. Two texts with similar meaning get vectors that are close in cosine space. Used for RAG (retrieval-augmented generation) over the 49GB UK government data moat and any document collection.

Models in the estate

Embedder	`BGE-M3` · 1024-dim · multilingual (100+ langs)
Re-ranker	`BGE-reranker` · cross-encoder, top-k refinement

Why we use it

BGE-M3 handles English + EU languages (French, German, Spanish) for compliance docs
Powers Q&A over Companies House, Land Registry, OS Open Names (49GB moat)
The 6 Hives (Security / Finance / Governance / Research / Operations / Creative) all use BGE-M3 for cross-source queries

Live invocation

# Embed + retrieve
python3 -c "
from sovereign_temple.retrieval import SovereignRAG
rag = SovereignRAG(embedding_model='bge-m3')
rag.index('/data/hive-data/governance/')
hits = rag.query('EU AI Act Article 50 transparency', top_k=5)
for hit in hits: print(hit.path, hit.score)
"

Live, indexes all 6 Hives

07 / Our Trained Neural Networks

CSOAI-OWNED · GOVERNANCE SIGNALS

Our 7 Trained NNs

Small specialist classifiers trained on real labelled data. Not a foundation model — deliberate narrow classifiers.

CSOAI own

What they are

Where the 8 model types above are general-purpose, these 7 are narrow specialist classifiers trained on labelled data CSOAI produces itself. They detect specific signals that broader models miss.

The full scorecard (honest)

creativity_assessment_nn	r² 0.91 · 350 samples · ✅ strong
care_pattern_analyzer	mae 0.037 · 600 samples · ✅ strong
relationship_evolution_nn	mae 0.071 · 500 samples · ✅ strong
care_validation_nn	mae 0.19 · 19 samples · ⚠️ tiny
partnership_detection_ml	mae 0.22 · 19 samples · ⚠️ tiny
threat_detection_nn	acc 0.45 · 33 samples · ❌ weak, retrain
dependency_detection_nn	acc 0.22 · 50 samples · ❌ weak, retrain

Why we use them

The 3 strong ones are deployed in production for governance signals
The 2 weak ones (threat / dependency) need retraining on real labelled data — honest flag
Each is narrow on purpose — they emit signals, not conversations
Combined into single governance verdict by the OOWM orchestration layer

Live invocation

# Call into the MEOK MCP gateway
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "sov3_run_governance_signals",
    "input": {"text": "...", "context": "..."},
    "models": ["care_pattern_analyzer", "creativity_assessment_nn"]
  }'
→ { care_pattern: 0.83, creativity: 0.71, signed_sigil: "..." }

3 models live · 2 need retraining

08 / Cloud Ensemble (Vote + Fallback)

⌁

VOTE / FALLBACK · API

Cloud Ensemble

External frontier models used as voters / fallbacks, never as primary route. API only — no sovereign-train data on them.

API · No train data

What it is

For tasks where local sovereign models aren't strong enough, we route to external frontier models (OpenAI, Anthropic Claude, GLM-5.2, Groq) — but never as the primary answer. They vote as part of a BFT ensemble, or act as fallback when local models decline. Critically: no user data is ever used to train them.

Models in the ensemble

Frontier (voter)	`OpenAI` · `Claude` · Anthropic API
Chinese cloud	`GLM-5.2` · sovereign-replaceable
Speed fallback	`Groq` · low-latency secondary

Why we use it

Adds diverse voter perspectives to BFT council decisions
Fallback when local sovereign model is overloaded (W38 honest finding)
GLM-5.2 used for Chinese-language compliance work (PIPL, AIGC reg.)
Rule: train data stays on sovereign side; API calls only

Live invocation

# Hybrid sovereign: local + 1 cloud voter
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "sov3_apply_routing",
    "category": "compliance",
    "voters": ["deepseek-r1:7b", "claude-sonnet-4"],
    "quorum": 2
  }'
→ returns BFT-signed verdict with both votes visible

Live as BFT voters