Sovereign Omniscient Vessel³. The brain that powers DEFONEOS. Mamba-2 state memory + 64-expert mixture + 33-node BFT governance. 222 tools. 3,000 tok/s. Zero foreign cloud.
AI tools across 8 categories: governance, compliance, creativity, memory, security, vision, voice, coordination
On-device inference throughput. No API calls to foreign servers.
22-of-33 Byzantine fault tolerant quorum for every critical decision
Compresses infinite context into a 16-dimensional state vector
SOV3 uses an Organic Open World Model (OOWM) sandwich — a Mamba-2 state-space model for infinite-length context compression, flanked by an attention-based Transformer for in-context reasoning, wrapped in a BFT council for governance. This is not a GPT wrapper. It is a sovereign cognition engine.
Every tool call, every BFT vote, every agent exchange emits a signed SIGIL receipt onto a tamper-evident hash chain. Full audit trail replayable by regulators. Zero-trust by design.
33 sovereign agents deliberate. 22-of-33 quorum required for critical actions. Tolerates 10 Byzantine (compromised) nodes. JSP 936 compliance assertion attached to every decision.
State-space model compresses infinite-length sensor data into a 16-dimensional state vector. Processes 198+ data streams simultaneously. O(1) memory per timestep — no sequence length limit.
The bind layer. Combines left-brain reasoning with right-brain perception into sovereign understanding. Routes queries to the optimal hemisphere based on task type.
Mixture-of-Experts with 64 specialised expert models. Only top-8 activated per query (sparse routing). Each expert specialised: code, legal, tactical, medical, logistics, cyber, intelligence, etc.
Model Context Protocol marketplace. 310+ servers, 1,700+ tools. DEFONEOS exposes 30 defense-specific MCPs. All callable from any agent, any framework, any language.
33 GCP VMs across 8 sovereign regions. All data stays in UK/EU. No CLOUD Act exposure. Quantum-safe key rotation (ML-DSA-65 + ML-KEM-768). Air-gap deployable.
SIGIL events stream at 1Hz into the Mamba-2 state vector. When the state shifts beyond a cosine similarity threshold, the system flags a "hunch" — a pre-conscious pattern detection. If 3+ consecutive states match, intuition is confirmed and surfaced to the BFT council. This gives DEFONEOS a 40-second lead time on emerging situations before they're consciously analysed.
SOV3 doesn't load one model for everything. It maintains 64 specialised experts and dynamically routes each query to the top-8 most relevant. This is sparse activation — only 12.5% of the network fires per query, delivering both speed and specialisation.
| Expert Domain | Specialisation | Activation Frequency |
|---|---|---|
| Tactical Analysis | Battlefield geometry, force disposition, OODA loop acceleration | 34% |
| Legal & Compliance | JSP 936, JSP 440, EU AI Act, GDPR, Article 50 | 28% |
| Intelligence Fusion | OSINT, SIGINT, GEOINT cross-correlation | 22% |
| Logistics & Supply | Supply chain optimisation, resource allocation, MEDLEVAC routing | 18% |
| Cyber Warfare | Threat detection, STRIDE analysis, PQC migration | 16% |
| Code & Systems | Full-stack engineering, MCP building, API design | 31% |
| Creative & Strategy | Bisociative thinking, cross-domain synthesis, doctrine drafting | 14% |
| Medical & CBRN | Casualty triage, chemical/biological threat assessment | 8% |
| Maritime | AIS tracking, naval domain awareness, sea-lane security | 11% |
| Aerospace | ADS-B tracking, ISR coordination, airspace deconfliction | 13% |
| Civil Services | Government data integration, census, Companies House | 19% |
| Governance | BFT deliberation, charter enforcement, audit trail | 24% |
Beyond the 64 MoE experts, SOV3 wraps 8 best-in-class models selected through competitive benchmarking (the BIG BRAIM protocol). Each is the category winner for its domain:
Every decision DEFONEOS makes is recorded as a SIGIL — a cryptographically signed event on a hash-chained ledger. This is not logging. This is proof. Regulators can replay any decision, verify the signature, and confirm the chain hasn't been tampered with.
| SIGIL Field | Description | Example |
|---|---|---|
| op | Operation type (P=propose, V=vote, M=message, Q=query, C=complete, H=handoff, S=sign, A=assert) | V |
| actor | Which agent emitted the SIGIL | sov3.bft.node12 |
| target | Intended recipient | council |
| payload | The actual content/command | deploy_counterdrone(dst=grid_ref) |
| digest | SHA-256 hash of payload | a3f8e2c1... |
| prev_hash | Hash of previous SIGIL (chain link) | 7b9d0f4e... |
| signature | Ed25519 signature over (digest + prev_hash) | ed25519:9c2a... |
| timestamp | ISO 8601 UTC | 2026-07-05T11:15:00Z |
Every SIGIL can carry an L3 compliance assertion — a signed attestation that the action complies with a specific framework article. This creates a three-layer audit pattern:
Ed25519 public key registered before any action. Agent identity is cryptographically provable.
Every action logged with hash chain to previous. Impossible to insert or delete entries without breaking the chain.
Each execution tagged with framework compliance (JSP 936 Art X, EU AI Act Art Y, GDPR Art Z). Signed attestation.
Critical decisions in DEFONEOS never come from a single agent. They are deliberated by a 33-node Byzantine Fault Tolerant council. 22-of-33 must agree. The system tolerates up to 10 compromised nodes.
| Parameter | Value |
|---|---|
| Total Nodes | 33 |
| Quorum Required | 22 (2/3 supermajority) |
| Max Byzantine Tolerance | 10 nodes (≈30%) |
| Deliberation Timeout | 30 seconds (fast path) |
| Vote Types | For / Against / Abstain |
| Signature Scheme | Ed25519 (one signature per node per vote) |
| Transparency | All votes SIGIL-emitted, publicly auditable |
Ensures all decisions comply with the SOV3 charter and JSP 936
Validates signatures, checks for injection, enforces red lines
Guards data sovereignty — blocks any action that would exfiltrate
Evaluates dual-use concerns, civilian impact, proportionality
Optimises for operational outcome within constraints
Correlates multi-source intelligence for situational awareness
SOV3 exposes 222+ tools via MCP (Model Context Protocol). Any agent, any framework, any language can call them. Here are the key categories:
| Category | Tools | Key Functions |
|---|---|---|
| Governance | bft_vote, council_propose, charter_query, sigil_emit, sigil_transcript | BFT deliberation, charter enforcement, audit trail |
| Compliance | article50_passport_issue, compliance_check, orgkernel_assert, jsp936_generate | EU AI Act, JSP 936, GDPR audit assertions |
| Memory | record_memory, query_memories, quantum_memory_search, temporal_chain | Episodic memory with care-weighting, VQE scoring |
| Cognition | oowm_think, bridge_think, zamba_ask, intuition_ingest, intuition_status | Left-right brain synthesis, Mamba-2 SSM, Zamba hybrid |
| Creativity | assess_creativity, find_bisociations, compute_novelty, apply_resonance | MAP-Elites archive, Koestler bisociation, 47 traditions |
| Federation | mcp_federation_search, mcp_federation_call, vault_search, federated_rag | 310+ MCP servers, 1,232 indexed files, E2E RAG |
| Security | detect_threats, dorado_detect, soc_bot_detector, key_rotation | Foreign-access detection, PQC key rotation, SOC tooling |
| Coordination | coord_submit_task, delegate_task, swarm_orchestrate, swarm_review | Multi-agent task delegation, Aegis reviewer gate |
| Config | Models Active | Latency | Use Case |
|---|---|---|---|
local_only | 4 on-device models (7GB total) | <100ms | Edge deploy, air-gap, Jetson |
balanced | 8 models, BFT off | ~500ms | Development, testing |
power | 14 models + Mamba-2 | ~2s | Full intelligence analysis |
council | 14 models + 33-node BFT | ~5s | Critical decisions, audit-grade |