🐉 SOV3³ + OOWM — Operator Runbook

The MoE is the default for deploy, sovereign, and defence task categories. The SOV3 router sends those tasks here unless overridden.

# Confirm route
curl -s -X POST http://localhost:3101/mcp \
  -d '{"tool":"sov3_oowm_routing","category":"deploy"}'
→ {"brain":"moe_large_online","model":"qwen3:30b-a3b","mindset":"decisive"}

# Direct Ollama call (bypasses SOV3 routing)
curl -s -X POST http://localhost:11434/api/generate \
  -d '{
    "model": "qwen3:30b-a3b",
    "prompt": "Summarise Article 50 of the EU AI Act in 5 bullets",
    "stream": false,
    "options": {"temperature": 0.3, "num_ctx": 8192}
  }' | jq .response

# Same task, routed through sovereign sandwich
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "sov3_apply_routing",
    "category": "deploy",
    "query": "Summarise Article 50 of the EU AI Act in 5 bullets"
  }'
→ Auto-routes, signs response with Ed25519, returns sovereign envelope

• Router needs to decide FAST which downstream model to use
• Resource-constrained Mac (16/32GB)
• Always-on classifier (junk/topic/route) — needs sub-100ms latency
• The fallback when qwen3:30b-a3b is unavailable

# Fast route call
curl -s http://localhost:11434/api/generate \
  -d '{"model": "qwen2.5:3b", "prompt": "Classify: GDPR-compliance / EU-AI-Act / other"}'

MoM is the default for world, physical, sovtown, companion — anything with image/sensor/spatial input.

# Route via SOV3
curl -s -X POST http://localhost:3101/mcp \
  -d '{"tool":"sov3_oowm_routing","category":"world"}'
→ {"brain":"mom_large_offline","model":"moondream:latest","mindset":"curious"}

# Direct vision call
python3 -c "
import base64, requests
img = base64.b64encode(open('/tmp/test.jpg','rb').read()).decode()
r = requests.post('http://localhost:11434/api/generate', json={
    'model': 'moondream',
    'prompt': 'What is in this image?',
    'images': [img]
})
print(r.json()['response'])
"

# Via mom_perceive MCP tool — returns structured + SIGIL
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "mom_perceive",
    "image_b64": "<base64>",
    "modality_hint": "objects",
    "return_format": "structured"
  }'
→ { objects: [...], confidence: 0.85, signed_sigil: "..." }

• Chart / table / diagram understanding (better than moondream)
• OCR of multi-language documents
• Spatial reasoning (where things are, not just what's there)

# Same shape as moondream, different model name
curl -s http://localhost:11434/api/generate \
  -d '{"model": "zamba-large", "prompt": "Where is the legend?", "images": ["<b64>"]}'

# For chart parsing → structured data
curl -s http://localhost:11434/api/generate \
  -d '{
    "model": "qwen-vl",
    "prompt": "Extract the data table from this chart as JSON",
    "images": ["<b64>"]
  }'

Mamba-2 isn't called by user routes — it's called by the OOWM engine itself to maintain world state. You can't POST /api/generate it directly. You interact through the OOWM.

# Inside sovereign_temple on the Mac
from sovereign_temple.sov3_oowm import OOWM
oowm = OOWM(brain="sovereign-intuition")

# Read the world state as a 16-dim vector
state = oowm.read_state()
print(state.shape)  → torch.Size([16])
print(state.norm())  → tensor(1.2341)   # bounded, never explodes

# Each new observation nudges state
oowm.update_state("user_asked_about_article_50")
oowm.update_state("user_passed_gdpr_consent")
oowm.update_state("user_requested_pro_tier")

# State is signed + appended to log (Ed25519)
assert oowm.last_receipt.sig is not None

DeepSeek-R1 is the default for compliance, sov3, ethics, reasoning — anywhere step-by-step logic matters.

# Route verification
curl -s -X POST http://localhost:3101/mcp \
  -d '{"tool":"sov3_oowm_routing","category":"compliance"}'
→ {"brain":"hybrid_sovereign","model":"deepseek-r1:7b","mindset":"wise"}

# Long-running reasoning call (3-15s thinking + final answer)
curl -s http://localhost:11434/api/generate \
  -d '{
    "model": "deepseek-r1:7b",
    "prompt": "Does this AI system violate Art 14? Steps: 1) read description 2) enumerate 4 human-oversight requirements 3) check each 4) emit verdict",
    "stream": false,
    "options": {"num_predict": 8192}
  }'
→ { thinking: "...step trace...", response: "VERDICT: REMEDIATE..." }

# Sovereign voice synthesis → wav
python3 -c "
from kokoro import synthesize
synthesize(
    text='This system complies with EU AI Act Article 50.',
    voice='bm_george',
    out_path='/tmp/compliance.wav'
)
print('OK')
"

# os.meok.ai voice button hits this endpoint
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "sov3_speak",
    "text": "Compliance check complete.",
    "voice": "bf_emma"
  }'
→ { audio_url: "/sovereign/voice/abc123.wav", sigil: "..." }

# Single text → 1024-dim vector
python3 -c "
from sovereign_temple.retrieval import SovereignEmbeddings
emb = SovereignEmbeddings(model='bge-m3')
vec = emb.encode('EU AI Act Article 50 requires watermark provenance')
print(vec.shape)  # → (1024,)
"

# Example: index the Governance Hive (30 sources, EU data corpus)
python3 -c "
from sovereign_temple.retrieval import SovereignRAG
rag = SovereignRAG(embedding_model='bge-m3')
rag.index('/data/hive-data/governance/')

# Query with re-ranking
hits = rag.query(
    'transparency obligations for AI-generated content',
    top_k=10,
    rerank_top_k=5
)
for hit in hits:
    print(f'{hit.score:.3f} {hit.path}')
"

→ 0.842 /data/hive-data/governance/eu_ai_act_art_50.md
   0.815 /data/hive-data/governance/eu_ai_act_art_12.md
   ...

# MEOK MCP exposes them all
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: "..." }

Cloud models participate as BFT voters (adds diverse perspective to council votes) and as fallback when local sovereign model is saturated. No user data is ever sent for training. Verify each integration with a data-handling attestation first.

# Same as compliance call, but with voters array
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "sov3_apply_routing",
    "category": "compliance",
    "query": "Score this AI system against Article 50",
    "voters": ["deepseek-r1:7b", "claude-sonnet-4"],
    "quorum": 2
  }'
→ BFT result with both votes, Ed25519-signed

# If local model is saturated, the voter takes primary
curl -s -X POST http://localhost:3101/mcp \
  -d '{
    "tool": "sov3_apply_routing",
    "category": "speed",
    "voters": ["qwen3:0.6b", "groq-llama"],
    "timeout_ms": 2000
  }'