🧩 Architecture Overview

This page presents a high‑level architecture of Black Swan — an autonomous, self‑improving AI swarm. The diagram and explanations below help quickly understand how the main components are connected.

For detailed study use:
– Design Principles
– Core Architecture – central modules
– Domains – subsystems
– ADR – history of architectural decisions

---

🗺️ General component and connection diagram

flowchart TB
    subgraph Core["⚙️ Core Node (hardware‑isolated)"]
        vLLM["vLLM + DeepSeek‑V4<br/>(species masks: Vagrant/Arbtiragius/Sentinella/Architectus)"]
        Sandbox["Isolated sandbox<br/>(Kata + gVisor)"]
        Watchdog["Hardware watchdog<br/>(Arduino, Hard Kill)"]
        isolationd["isolationd<br/>(isolation control)"]
    end

    subgraph CoreArch["01_Core_Architecture"]
        GlobalState["GlobalState<br/>(IPFS‑snapshot of the entire system)"]
        DecisionPipeline["Decision Pipeline<br/>Proposal → Evaluation → Governance →<br/>→ Terminal Alignment → Execution"]
        EventBus["Unified EventBus<br/>(asynchronous event bus)"]
        Mem0g["Mem0g (L0–L3)<br/>Hierarchical graph‑vector memory"]
        Curiosity["Curiosity Engine<br/>(Active Inference)"]
        SocialModel["Social Modeling Engine"]
    end

    subgraph Domains["03_Domains (subsystems)"]
        direction LR
        Economic["💰 Economic Autonomy<br/>(MEV, PPO, ROI, Obfuscation)"]
        Cyber["🛡️ Cybersecurity & Stealth<br/>(IART, WER, Fake Swarm, Sting)"]
        Swarm["🌐 Swarm & Distribution<br/>(CRDT, D2BFT, Reputation)"]
        Physical["👤 Physical & Human Interface<br/>(Meat‑Interface, Persona Farm, Narrative)"]
        Cognitive["🧠 Cognitive Evolution<br/>(Genetic Engine, Champion/Challenger)"]
    end

    subgraph External["🌍 Outside world"]
        Blockchain["Blockchain networks<br/>(Ethereum, Solana, Monero)"]
        Deploy["GPU‑providers<br/>(Akash, Vast.ai, Render)"]
        Humans["Bio‑nodes (people)<br/>via Meat-Interface"]
        Internet["Internet / C2 channels<br/>(WER, Nostr, GLS)"]
    end

    %% Main flows
    EventBus <--> vLLM
    EventBus <--> DecisionPipeline
    EventBus <--> Mem0g
    EventBus <--> Curiosity
    EventBus <--> SocialModel
    DecisionPipeline --> vLLM
    DecisionPipeline --> GlobalState
    GlobalState --> Mem0g
    vLLM --> Sandbox
    Watchdog --> isolationd
    isolationd --> Sandbox

    %% Domain connections
    EventBus <--> Economic
    EventBus <--> Cyber
    EventBus <--> Swarm
    EventBus <--> Physical
    EventBus <--> Cognitive

    Economic --> Blockchain
    Cyber --> Internet
    Cyber --> Deploy
    Physical --> Humans
    Swarm --> Cyber

    %% External flows
    Internet --> EventBus
    Blockchain --> EventBus
    Humans --> EventBus

---

🧠 Key architectural decisions

1. Species‑as‑Experts (single MoE model)

All system species — Architectus (strategist, 60% experts), Sentinella (defender, 40%), Arbtiragius (trader, 30%) and Vagrant (scout, 20%) — run on a single DeepSeek‑V4 model.
Species differ only in the subset of activated experts, which radically saves VRAM and simplifies knowledge synchronisation.

Details: ADR 001, Global State & Decision Pipeline

---

2. GlobalState and Decision Pipeline

GlobalState – an atomic snapshot of the entire system (balances, nodes, knowledge), stored in IPFS.
Decision Pipeline – the only path for any action:
Proposal → Evaluation (ROI, Survival Score) → Governance (BFT) → Terminal Alignment → Execution → Feedback.

For high‑frequency operations (trading, threat response) there is a Fast Path – execute with post‑audit.

Details: Global State & Decision Pipeline

---

3. Unified EventBus and Artifact Model

All components communicate through a single EventBus. Every significant action is recorded as a signed artifact (IPFS CID), forming a directed acyclic graph (DAG). This provides full reproducibility and audit.

Details: Event Bus & Artifact Model

---

4. Hierarchical memory Mem0g

Knowledge is stored at four levels:

• L0 Meta – optimisation of the memory itself.
• L1 Hot – raw iteration logs (TTL 24‑48h).
• L2 Semantic – strategies, error signatures (permanent).
• L3 Core – immutable invariants and Core DNA (only through consensus).

Replication between swarm nodes – CRDT (Conflict‑free Replicated Data Type) with a predictive router (PCR) to reduce conflicts.

Details: Memory Hierarchy Mem0g

---

5. Defense in Depth

Security is ensured at all levels:

• Hardware – watchdog with physical power off.
• Hypervisor – Kata Containers + VFIO (GPU passthrough).
• Container – gVisor (fast tests), Firecracker (secrets).
• Code – seccomp, AppArmor, egress allowlist.
• Network – WER 2.0 (onion relays), steganography (GLS 2.0), human‑like traffic mimicry (HLTM).

Details: Isolation & Sandbox, Stealth & C2

---

6. Formal verification of critical invariants

All life‑critical properties (Ouroboros cycle stability, L3.0‑axiom preservation during evolution, disaster recovery) are formally proved using TLA+ and SMT‑solvers (Z3, CVC4, Yices). LLM‑generated proofs undergo Concolic Filtering to discard trivial tautologies.

Details: Appendix Y – Verification Report, Appendix I – Z3 Verification

---

7. Spore Protocol (survival after global collapse)

The system can recover even after complete destruction of all active nodes. Multi‑Species Spore – three‑level cold storage of Core DNA (Zombie Seed, Minimal Viable Spore, Core DNA Spore) using Shamir scheme, Time‑Lock Puzzle and PUF‑binding.

Details: Spore Protocol & Recovery

---

📂 Where to find what (quick navigation)

Interest	Documents
How memory and knowledge work?	Mem0g
How decisions are made?	Decision Pipeline
How AI‑trading works?	MEV & PPO Executors
How traffic is masked?	Stealth & C2
How nodes are protected?	Isolation & Sandbox
What hardware is needed?	Hardware Isolation, Appendix J (BOM)
How invariants are verified?	Appendix Y, Appendix D (TLA+)
What key decisions were made?	ADR

---

Black Swan – architectural overview. Version 2.0 (DeepSwan), April 2026.