EU AI Act Article 50 — 20 days to seal | Get passport

DEFONEOS FAQ v2: Expanded Technical Questions

Q: What is the underlying architecture of DEFONEOS? ▼

DEFONEOS is built on a 5-layer Sovereign Substrate architecture:

  1. Layer 0: Core Protocols (MCP, A2A, x402, DID, JWT, IBC, OIDC)
  2. Layer 1: Identity & Attestation (W3C DIDs, Ed25519-signed SIGILs, BFT Council for governance)
  3. Layer 2: Execution & Orchestration (SOV3 Hive, Federated MCPs, Organic Learning Model (OLM) for routing)
  4. Layer 3: Cognitive & Analytical (BIG BRAIM multi-model ensemble, Mamba-2 SSM for long context, MoE for reasoning, MOM for perception)
  5. Layer 4: Physical Actuation & Digital Twin (iOK Farm demo, Cesium 3D Globe, Open Hands OS for human-AI interaction)

The entire stack is designed for air-gapped deployment, UK sovereignty, and auditability via its hash-chained SIGIL ledger.

Q: How does DEFONEOS ensure data sovereignty and control? ▼

Data sovereignty is fundamental to DEFONEOS, implemented through several mechanisms:

  • Ed25519 Signatures: Every inter-agent message (SIGIL) and critical action is cryptographically signed, creating an immutable, auditable chain of custody.
  • Compartment Doctrine: Strict separation of code and IP for `meok-defoneos` (builds), `csoai-defoneos` (certifies), and `dagon` (legacy/NDA-only) ensures no cross-contamination.
  • Geo-fencing & Multi-region Deployment: Data can be isolated to specific sovereign regions (e.g., UK, AUKUS) with ZK-SNARK proofs of data residency.
  • Air-gapped Deployment: The OS is designed to function entirely offline, preventing unauthorized exfiltration.
  • OrgKernel 3-Layer Audit: Identity (L1), Execution (L2), and Compliance (L3) are logged and verified on-chain, providing granular accountability.
Q: What is a "SIGIL" and how is it used in DEFONEOS? ▼

A SIGIL (Signed Inter-Agent Global Interchange Ledger) is the atomic unit of communication and action within the DEFONEOS ecosystem. It's a cryptographically signed, hash-chained entry on an immutable ledger.

Key uses:

  • Audit Trail: Provides a transparent, unforgeable record of all agent interactions, decisions, and data flows for regulatory compliance (EU AI Act, JSP 936).
  • Inter-Agent Communication: Agents (like JEEVES, JARVIS, Kimi) communicate by emitting SIGILs, ensuring all actions are traceable and attributable.
  • Training & Learning: The Organic Learning Model (OLM) learns from the patterns and outcomes recorded in SIGILs to improve its routing and decision-making over time.
  • Consensus & Governance: Critical actions, like BFT Council votes or model deployments, are recorded as SIGILs.

Example SIGIL line:

H|opus|sov3|review the Q3 plan|aeef2e7e54809755|Ed25519_SIG_DATA
Q: How does DEFONEOS handle AI model provenance and integrity? ▼

Model provenance and integrity are critical. DEFONEOS implements:

  • C2PA Integration: Every generated output (text, image, audio, video) is watermarked and signed with C2PA metadata, allowing verifiable tracing back to the originating model and input.
  • Model Weight Hashing & Integrity Scans: Before deployment, model weights are hashed and continuously scanned for tampering or poisoning attempts.
  • BFT Council Approval: Deployment of new or updated models requires a Byzantine Fault Tolerant (BFT) council vote (quorum 23/33 agents) to ensure multi-stakeholder approval.
  • Supply Chain Risk Management: MCP manifests include detailed dependency trees, allowing for proactive vulnerability identification.
  • Article 50 Passports: For EU AI Act compliance, each AI-generated content piece is issued an Article 50 watermarking passport, detailing its origin and watermarking status.
Q: What is the "BIG BRAIM" and how does it contribute to DEFONEOS? ▼

The "BIG BRAIM" is SOV3's sovereign multi-model ensemble, wrapping 8 category-winning LLMs and AI models into a single, cohesive cognitive unit. It acts as a universal router, intelligently selecting the best model for a given task.

Categories include:

  • Coding: (e.g., `falcon3-code`) for software development and analysis.
  • Reasoning: (e.g., `deepseek-r1`) for complex problem-solving.
  • Long Context: (e.g., `mamba-2`) for processing extensive documents and histories.
  • Multilingual: (e.g., `qwen2.5`) for diverse language handling.
  • Edge: (e.g., `llama3.1-4bit`) for optimized on-device performance.
  • TTS & Embedding: Specialized models for text-to-speech and vector embeddings.
  • Router: An internal OLM-trained router that intelligently directs queries.

This allows DEFONEOS to leverage the strengths of various models while maintaining a unified, sovereign control plane, with every invocation recorded via SIGILs.

Q: How does DEFONEOS enable "Human-in-the-Loop" decision making? ▼

Human-in-the-Loop (HITL) is paramount for ethical and responsible AI deployment, especially in defence. DEFONEOS integrates HITL at multiple levels:

  • Red Lines Enforcement: Immutable red lines (e.g., no kinetic targeting, no personal surveillance) are hard-coded into the system and enforced by the BFT Council, preventing autonomous action in sensitive areas.
  • Operator Overlays (Open Hands OS): The Open Hands OS provides intuitive user interfaces (TUI, native mobile apps, 3D globes) that give human operators clear visibility and control over AI actions.
  • Plausibility Checks: AI suggestions are always subject to human plausibility checks, with mechanisms to flag and review anomalous outputs.
  • Consent & Digital Twins: For i-character generation (digital twins), explicit user consent is required, ensuring autonomous agents respect human autonomy.
  • BFT Veto Mechanisms: The BFT Council can veto any proposed AI action that deviates from ethical guidelines or operational protocols.