Files

ruv cd5943df23 Merge commit 'd803bfe2b1fe7f5e219e50ac20d6801a0a58ac75' as 'vendor/ruvector'

2026-02-28 14:39:40 -05:00

51 KiB

Raw Blame History

@ruvector/edge-net

Collective AI Computing Network - Share, Contribute, Compute Together

A distributed computing platform that enables collective resource sharing for AI workloads. Contributors share idle compute resources, earning participation units (rUv) that can be used to access the network's collective AI computing power.

┌─────────────────────────────────────────────────────────────────────────────┐
│              EDGE-NET: COLLECTIVE AI COMPUTING NETWORK                      │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   ┌─────────────┐       ┌─────────────┐       ┌─────────────┐              │
│   │  Your       │       │  Collective │       │  AI Tasks   │              │
│   │  Browser    │◄─────►│  Network    │◄─────►│  Completed  │              │
│   │  (Idle CPU) │  P2P  │  (1000s)    │       │  for You    │              │
│   └─────────────┘       └─────────────┘       └─────────────┘              │
│         │                     │                     │                       │
│         ▼                     ▼                     ▼                       │
│   ┌─────────────┐       ┌─────────────┐       ┌─────────────┐              │
│   │  Contribute │       │  Earn rUv   │       │  Use rUv    │              │
│   │  Compute    │  ───► │  Units      │  ───► │  for AI     │              │
│   │  When Idle  │       │  (Credits)  │       │  Workloads  │              │
│   └─────────────┘       └─────────────┘       └─────────────┘              │
│                                                                             │
│   Vector Search │ Embeddings │ Semantic Match │ Encryption │ Compression   │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

What is Edge-Net?
Key Features
Quick Start
How It Works
AI Computing Tasks
Pi-Key Identity System
Self-Optimization
Tutorials
API Reference
Development
Exotic AI Capabilities
Core Architecture & Capabilities
Self-Learning Hooks & MCP Integration

What is Edge-Net?

Edge-net creates a collective computing network where participants share idle browser resources to power distributed AI workloads. Think of it as a cooperative where:

You Contribute - Share unused CPU cycles when browsing
You Earn - Accumulate rUv (Resource Utility Vouchers) based on contribution
You Use - Spend rUv to run AI tasks across the collective network
Network Grows - More participants = more collective computing power

Why Collective AI Computing?

Traditional AI Computing	Collective Edge-Net
Expensive GPU servers	Free idle browser CPUs
Centralized data centers	Distributed global network
Pay-per-use pricing	Contribution-based access
Single point of failure	Resilient P2P mesh
Limited by your hardware	Scale with the collective

Core Principles

Principle	Description
Collectibility	Resources are pooled and shared fairly
Contribution	Earn by giving, spend by using
Self-Sustaining	Network operates without central control
Privacy-First	Pi-Key cryptographic identity system
Adaptive	Q-learning security protects the collective

Key Features

Feature	Benefit
Idle CPU Utilization	Use resources that would otherwise be wasted
Browser-Based	No installation, runs in any modern browser
Adjustable Contribution	Control how much you share (10-50% CPU)
Battery Aware	Automatically reduces on battery power
Fair Distribution	Work routed based on capability matching

AI Computing Capabilities

Edge-net provides a complete AI stack that runs entirely in your browser. Each component is designed to be lightweight, fast, and work without a central server.

┌─────────────────────────────────────────────────────────────────────────────┐
│                        AI INTELLIGENCE STACK                                 │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│  ┌─────────────────────────────────────────────────────────────────────┐   │
│  │                    MicroLoRA Adapter Pool (from ruvLLM)              │   │
│  │  • LRU-managed pool (16 slots) • Rank 1-16 adaptation                │   │
│  │  • <50µs rank-1 forward        • 2,236+ ops/sec with batch 32        │   │
│  │  • 4-bit/8-bit quantization    • P2P shareable adapters              │   │
│  └─────────────────────────────────────────────────────────────────────┘   │
│                                                                             │
│  ┌─────────────────────────────────────────────────────────────────────┐   │
│  │                    SONA - Self-Optimizing Neural Architecture         │   │
│  │  • Instant Loop: Per-request MicroLoRA adaptation                    │   │
│  │  • Background Loop: Hourly K-means consolidation                     │   │
│  │  • Deep Loop: Weekly EWC++ consolidation (catastrophic forgetting)   │   │
│  └─────────────────────────────────────────────────────────────────────┘   │
│                                                                             │
│  ┌──────────────────────┐  ┌──────────────────────┐  ┌─────────────────┐  │
│  │   HNSW Vector Index  │  │  Federated Learning  │  │ ReasoningBank   │  │
│  │   • 150x faster      │  │  • TopK Sparsify 90% │  │ • Trajectories  │  │
│  │   • O(log N) search  │  │  • Byzantine tolerant│  │ • Pattern learn │  │
│  │   • Incremental P2P  │  │  • Diff privacy      │  │ • 87x energy    │  │
│  └──────────────────────┘  └──────────────────────┘  └─────────────────┘  │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

Core AI Tasks

Task Type	Use Case	How It Works
Vector Search	Find similar items	HNSW index with 150x speedup
Embeddings	Text understanding	Generate semantic vectors
Semantic Match	Intent detection	Classify meaning
LoRA Inference	Task adaptation	MicroLoRA <100µs forward
Pattern Learning	Self-optimization	ReasoningBank trajectories

MicroLoRA Adapter System

What it does: Lets the network specialize for different tasks without retraining the whole model. Think of it like having 16 expert "hats" the AI can quickly swap between - one for searching, one for encryption, one for routing, etc.

Ported from ruvLLM with enhancements for distributed compute:

Feature	Specification	Performance
Rank Support	1-16	Rank-1: <50µs, Rank-2: <100µs
Pool Size	16 concurrent adapters	LRU eviction policy
Quantization	4-bit, 8-bit	75% memory reduction
Batch Size	32 (optimal)	2,236+ ops/sec
Task Types	VectorSearch, Embedding, Inference, Crypto, Routing	Auto-routing

Why it matters: Traditional AI models are "one size fits all." MicroLoRA lets each node become a specialist for specific tasks in under 100 microseconds - faster than a blink.

SONA: Self-Optimizing Neural Architecture

What it does: The network teaches itself to get better over time using three learning speeds - instant reactions, daily improvements, and long-term memory. Like how your brain handles reflexes, daily learning, and permanent memories differently.

Three-temporal-loop continuous learning system:

Loop	Interval	Mechanism	Purpose
Instant	Per-request	MicroLoRA rank-2	Immediate adaptation
Background	Hourly	K-means clustering	Pattern consolidation
Deep	Weekly	EWC++ (λ=2000)	Prevent catastrophic forgetting

Why it matters: Most AI systems forget old knowledge when learning new things ("catastrophic forgetting"). SONA's three-loop design lets the network learn continuously without losing what it already knows.

HNSW Vector Index

What it does: Finds similar items incredibly fast by organizing data like a multi-level highway system. Instead of checking every item (like walking door-to-door), it takes smart shortcuts to find what you need 150x faster.

Parameter	Default	Description
M	32	Max connections per node
M_max_0	64	Max connections at layer 0
ef_construction	200	Build-time beam width
ef_search	64	Search-time beam width
Performance	150x	Speedup vs linear scan

Why it matters: When searching millions of vectors, naive search takes seconds. HNSW takes milliseconds - essential for real-time AI responses.

Federated Learning

What it does: Nodes teach each other without sharing their private data. Each node trains on its own data, then shares only the "lessons learned" (gradients) - like students sharing study notes instead of copying each other's homework.

P2P gradient gossip without central coordinator:

Feature	Mechanism	Benefit
TopK Sparsification	90% compression	Only share the most important updates
Rep-Weighted FedAvg	Reputation scoring	Trusted nodes have more influence
Byzantine Tolerance	Outlier detection, clipping	Ignore malicious or broken nodes
Differential Privacy	Noise injection	Mathematically guaranteed privacy
Gossip Protocol	Eventually consistent	Works even if some nodes go offline

Why it matters: Traditional AI training requires sending all your data to a central server. Federated learning keeps your data local while still benefiting from collective intelligence.

ReasoningBank & Learning Intelligence

What it does: The network's "memory system" that remembers what worked and what didn't. Like keeping a journal of successful strategies that any node can learn from.

Component	What It Does	Why It's Fast
ReasoningBank	Stores successful task patterns	Semantic search for quick recall
Pattern Extractor	Groups similar experiences together	K-means finds common patterns
Multi-Head Attention	Decides which node handles each task	Parallel evaluation of options
Spike-Driven Attention	Ultra-low-power decision making	87x more energy efficient

Why it matters: Without memory, the network would repeat the same mistakes. ReasoningBank lets nodes learn from each other's successes and failures across the entire collective.

Pi-Key Identity System

Ultra-compact cryptographic identity using mathematical constants:

Key Type	Size	Purpose
π (Pi-Key)	40 bytes	Your permanent identity
e (Session)	34 bytes	Temporary encrypted sessions
φ (Genesis)	21 bytes	Network origin markers

Self-Optimizing Network

Automatic Task Routing - Work goes to best-suited nodes
Topology Optimization - Network self-organizes for efficiency
Q-Learning Security - Learns to defend against threats
Economic Balance - Self-sustaining resource economy

Quick Start

1. Add to Your Website

<script type="module">
  import init, { EdgeNetNode, EdgeNetConfig } from '@ruvector/edge-net';

  async function joinCollective() {
    await init();

    // Join the collective with your site ID
    const node = new EdgeNetConfig('my-website')
      .cpuLimit(0.3)          // Contribute 30% CPU when idle
      .memoryLimit(256 * 1024 * 1024)  // 256MB max
      .respectBattery(true)   // Reduce on battery
      .build();

    // Start contributing to the collective
    node.start();

    // Monitor your participation
    setInterval(() => {
      console.log(`Contributed: ${node.ruvBalance()} rUv`);
      console.log(`Tasks completed: ${node.getStats().tasks_completed}`);
    }, 10000);
  }

  joinCollective();
</script>

2. Use the Collective's AI Power

// Submit an AI task to the collective
const result = await node.submitTask('vector_search', {
  query: embeddings,
  k: 10,
  index: 'shared-knowledge-base'
}, 5);  // Spend up to 5 rUv

console.log('Similar items:', result);

3. Monitor Your Contribution

// Check your standing in the collective
const stats = node.getStats();
console.log(`
  rUv Earned: ${stats.ruv_earned}
  rUv Spent: ${stats.ruv_spent}
  Net Balance: ${stats.ruv_earned - stats.ruv_spent}
  Tasks Completed: ${stats.tasks_completed}
  Reputation: ${(stats.reputation * 100).toFixed(1)}%
`);

How It Works

The Contribution Cycle

┌─────────────────────────────────────────────────────────────────────────────┐
│                        CONTRIBUTION CYCLE                                    │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   1. CONTRIBUTE          2. EARN              3. USE                        │
│   ┌─────────────┐       ┌─────────────┐       ┌─────────────┐              │
│   │   Browser   │       │    rUv      │       │  AI Tasks   │              │
│   │   detects   │  ───► │   credited  │  ───► │  submitted  │              │
│   │   idle time │       │   to you    │       │  to network │              │
│   └─────────────┘       └─────────────┘       └─────────────┘              │
│         │                     │                     │                       │
│         ▼                     ▼                     ▼                       │
│   ┌─────────────┐       ┌─────────────┐       ┌─────────────┐              │
│   │  Process    │       │  10x boost  │       │  Results    │              │
│   │  incoming   │       │  for early  │       │  returned   │              │
│   │  tasks      │       │  adopters   │       │  to you     │              │
│   └─────────────┘       └─────────────┘       └─────────────┘              │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

Network Growth Phases

The collective grows through natural phases:

Phase	Size	Your Benefit
Genesis	0-10K nodes	10x rUv multiplier (early adopter bonus)
Growth	10K-50K	Multiplier decreases, network strengthens
Maturation	50K-100K	Stable economy, high reliability
Independence	100K+	Self-sustaining, maximum collective power

Fair Resource Allocation

// The network automatically optimizes task distribution
const health = JSON.parse(node.getEconomicHealth());

console.log(`
  Resource Velocity: ${health.velocity}      // How fast resources circulate
  Utilization: ${health.utilization}         // Network capacity used
  Growth Rate: ${health.growth}              // Network expansion
  Stability: ${health.stability}             // Economic equilibrium
`);

AI Computing Tasks

Vector Search (Distributed Similarity)

Find similar items across the collective's distributed index:

// Search for similar documents
const similar = await node.submitTask('vector_search', {
  query: [0.1, 0.2, 0.3, ...],  // Your query vector
  k: 10,                         // Top 10 results
  index: 'shared-docs'           // Distributed index name
}, 3);  // Max 3 rUv

// Results from across the network
similar.forEach(item => {
  console.log(`Score: ${item.score}, ID: ${item.id}`);
});

Embedding Generation

Generate semantic embeddings using collective compute:

// Generate embeddings for text
const embeddings = await node.submitTask('embedding', {
  text: 'Your text to embed',
  model: 'sentence-transformer'
}, 2);

console.log('Embedding vector:', embeddings);

Semantic Matching

Classify intent or meaning:

// Classify text intent
const intent = await node.submitTask('semantic_match', {
  text: 'I want to cancel my subscription',
  categories: ['billing', 'support', 'sales', 'general']
}, 1);

console.log('Detected intent:', intent.category);

Secure Operations

Encrypt data across the network:

// Distributed encryption
const encrypted = await node.submitTask('encryption', {
  data: sensitiveData,
  operation: 'encrypt',
  key_id: 'my-shared-key'
}, 2);

Pi-Key Identity System

Your identity in the collective uses mathematical constants for key sizes:

Key Types

┌─────────────────────────────────────────────────────────────────────────────┐
│                        PI-KEY IDENTITY SYSTEM                               │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   π Pi-Key (Identity)     e Euler-Key (Session)    φ Phi-Key (Genesis)     │
│   ┌─────────────────┐     ┌───────────────┐        ┌───────────────┐       │
│   │   314 bits      │     │   271 bits    │        │   161 bits    │       │
│   │   = 40 bytes    │     │   = 34 bytes  │        │   = 21 bytes  │       │
│   │                 │     │               │        │               │       │
│   │   Your unique   │     │   Temporary   │        │   Origin      │       │
│   │   identity      │     │   sessions    │        │   markers     │       │
│   │   (permanent)   │     │   (encrypted) │        │   (network)   │       │
│   └─────────────────┘     └───────────────┘        └───────────────┘       │
│                                                                             │
│   Ed25519 Signing         AES-256-GCM              SHA-256 Derived         │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

Using Pi-Keys

import { PiKey, SessionKey, GenesisKey } from '@ruvector/edge-net';

// Create your permanent identity
const identity = new PiKey();
console.log(`Your ID: ${identity.getShortId()}`);  // π:a1b2c3d4...

// Sign data
const signature = identity.sign(data);
const valid = identity.verify(data, signature, identity.getPublicKey());

// Create encrypted backup
const backup = identity.createEncryptedBackup('my-password');

// Create temporary session
const session = SessionKey.create(identity, 3600);  // 1 hour
const encrypted = session.encrypt(sensitiveData);
const decrypted = session.decrypt(encrypted);

Security Architecture

Edge-net implements production-grade cryptographic security:

Cryptographic Primitives

Component	Algorithm	Purpose
Key Derivation	Argon2id (64MB, 3 iterations)	Memory-hard password hashing
Signing	Ed25519	Digital signatures (128-bit security)
Encryption	AES-256-GCM	Authenticated encryption
Hashing	SHA-256	Content hashing and verification

Identity Protection

// Password-protected key export with Argon2id + AES-256-GCM
let encrypted = identity.export_secret_key("strong_password")?;

// Secure memory cleanup (zeroize)
// All sensitive key material is automatically zeroed after use

Authority Verification

All resolution events require cryptographic proof:

// Ed25519 signature verification for authority decisions
let signature = ScopedAuthority::sign_resolution(&resolution, &context, &signing_key);
// Signature verified against registered authority public keys

Attack Resistance

The RAC (RuVector Adversarial Coherence) protocol defends against:

Attack	Defense
Sybil	Stake-weighted voting, witness path diversity
Eclipse	Context isolation, Merkle divergence detection
Byzantine	1/3 threshold, escalation tracking
Replay	Timestamp validation, duplicate detection
Double-spend	Conflict detection, quarantine system

Self-Optimization

The network continuously improves itself:

Automatic Task Routing

// Get optimal peers for your tasks
const peers = node.getOptimalPeers(5);

// Network learns from every interaction
node.recordTaskRouting('vector_search', 'peer-123', 45, true);

Fitness-Based Evolution

// High-performing nodes can replicate their config
if (node.shouldReplicate()) {
  const optimalConfig = node.getRecommendedConfig();
  // New nodes inherit successful configurations
}

// Track your contribution
const fitness = node.getNetworkFitness();  // 0.0 - 1.0

Q-Learning Security

The collective learns to defend itself:

// Run security audit
const audit = JSON.parse(node.runSecurityAudit());
console.log(`Security Score: ${audit.security_score}/10`);

// Defends against:
// - DDoS attacks
// - Sybil attacks
// - Byzantine behavior
// - Eclipse attacks
// - Replay attacks

Tutorials

Tutorial 1: Join the Collective

import init, { EdgeNetConfig } from '@ruvector/edge-net';

async function joinCollective() {
  await init();

  // Configure your contribution
  const node = new EdgeNetConfig('my-site')
    .cpuLimit(0.25)           // 25% CPU when idle
    .memoryLimit(128 * 1024 * 1024)  // 128MB
    .minIdleTime(5000)        // Wait 5s of idle
    .respectBattery(true)     // Reduce on battery
    .build();

  // Join the network
  node.start();

  // Check your status
  console.log('Joined collective!');
  console.log(`Node ID: ${node.nodeId()}`);
  console.log(`Multiplier: ${node.getMultiplier()}x`);

  return node;
}

Tutorial 2: Contribute and Earn

async function contributeAndEarn(node) {
  // Process tasks from the collective
  let tasksCompleted = 0;

  while (true) {
    // Check if we should work
    if (node.isIdle()) {
      // Process a task from the network
      const processed = await node.processNextTask();

      if (processed) {
        tasksCompleted++;
        const stats = node.getStats();
        console.log(`Completed ${tasksCompleted} tasks, earned ${stats.ruv_earned} rUv`);
      }
    }

    await new Promise(r => setTimeout(r, 1000));
  }
}

Tutorial 3: Use Collective AI Power

async function useCollectiveAI(node) {
  // Check your balance
  const balance = node.ruvBalance();
  console.log(`Available: ${balance} rUv`);

  // Submit AI tasks
  const tasks = [
    { type: 'vector_search', cost: 3 },
    { type: 'embedding', cost: 2 },
    { type: 'semantic_match', cost: 1 }
  ];

  for (const task of tasks) {
    if (balance >= task.cost) {
      console.log(`Running ${task.type}...`);
      const result = await node.submitTask(
        task.type,
        { data: 'sample' },
        task.cost
      );
      console.log(`Result: ${JSON.stringify(result)}`);
    }
  }
}

Tutorial 4: Monitor Network Health

async function monitorHealth(node) {
  setInterval(() => {
    // Your contribution
    const stats = node.getStats();
    console.log(`
      === Your Contribution ===
      Earned: ${stats.ruv_earned} rUv
      Spent: ${stats.ruv_spent} rUv
      Tasks: ${stats.tasks_completed}
      Reputation: ${(stats.reputation * 100).toFixed(1)}%
    `);

    // Network health
    const health = JSON.parse(node.getEconomicHealth());
    console.log(`
      === Network Health ===
      Velocity: ${health.velocity.toFixed(2)}
      Utilization: ${(health.utilization * 100).toFixed(1)}%
      Stability: ${health.stability.toFixed(2)}
    `);

    // Check sustainability
    const sustainable = node.isSelfSustaining(10000, 50000);
    console.log(`Self-sustaining: ${sustainable}`);

  }, 30000);
}

API Reference

Core Methods

Method	Description	Returns
`new EdgeNetNode(siteId)`	Join the collective	`EdgeNetNode`
`start()`	Begin contributing	`void`
`pause()` / `resume()`	Control contribution	`void`
`ruvBalance()`	Check your credits	`u64`
`submitTask(type, payload, maxCost)`	Use collective compute	`Promise<Result>`
`processNextTask()`	Process work for others	`Promise<bool>`

Identity Methods

Method	Description	Returns
`new PiKey()`	Generate identity	`PiKey`
`getIdentity()`	Get 40-byte identity	`Vec<u8>`
`sign(data)`	Sign data	`Vec<u8>`
`verify(data, sig, pubkey)`	Verify signature	`bool`
`createEncryptedBackup(password)`	Backup identity	`Vec<u8>`

Network Methods

Method	Description	Returns
`getNetworkFitness()`	Your contribution score	`f32`
`getOptimalPeers(count)`	Best nodes for tasks	`Vec<String>`
`getEconomicHealth()`	Network health metrics	`String (JSON)`
`isSelfSustaining(nodes, tasks)`	Check sustainability	`bool`

Development

Build

cd examples/edge-net
wasm-pack build --target web --out-dir pkg

Test

cargo test

Run Simulation

cd sim
npm install
npm run simulate

Exotic AI Capabilities

Edge-net can be enhanced with exotic AI WASM capabilities for advanced P2P coordination, self-learning, and distributed reasoning. Enable these features by building with the appropriate feature flags.

Available Feature Flags

Feature	Description	Dependencies
`exotic`	Time Crystal, NAO, Morphogenetic Networks	ruvector-exotic-wasm
`learning-enhanced`	MicroLoRA, BTSP, HDC, WTA, Global Workspace	ruvector-learning-wasm, ruvector-nervous-system-wasm
`economy-enhanced`	Enhanced CRDT credits	ruvector-economy-wasm
`exotic-full`	All exotic capabilities	All above

Time Crystal (P2P Synchronization)

Robust distributed coordination using discrete time crystal dynamics:

// Enable time crystal with 10 oscillators
node.enableTimeCrystal(10);

// Check synchronization level (0.0 - 1.0)
const sync = node.getTimeCrystalSync();
console.log(`P2P sync: ${(sync * 100).toFixed(1)}%`);

// Check if crystal is stable
if (node.isTimeCrystalStable()) {
  console.log('Network is synchronized!');
}

NAO (Neural Autonomous Organization)

Decentralized governance with stake-weighted quadratic voting:

// Enable NAO with 70% quorum requirement
node.enableNAO(0.7);

// Add peer nodes as members
node.addNAOMember('peer-123', 100);
node.addNAOMember('peer-456', 50);

// Propose and vote on network actions
const propId = node.proposeNAOAction('Increase task capacity');
node.voteNAOProposal(propId, 0.9);  // Vote with 90% weight

// Execute if quorum reached
if (node.executeNAOProposal(propId)) {
  console.log('Proposal executed!');
}

MicroLoRA (Per-Node Self-Learning)

Ultra-fast LoRA adaptation with <100us latency:

// Enable MicroLoRA with rank-2 adaptation
node.enableMicroLoRA(2);

// Adapt weights based on task feedback
const gradient = new Float32Array(128);
node.adaptMicroLoRA('vector_search', gradient);

// Apply adaptation to inputs
const input = new Float32Array(128);
const adapted = node.applyMicroLoRA('vector_search', input);

HDC (Hyperdimensional Computing)

10,000-bit binary hypervectors for distributed reasoning:

// Enable HDC memory
node.enableHDC();

// Store patterns for semantic operations
node.storeHDCPattern('concept_a');
node.storeHDCPattern('concept_b');

WTA (Winner-Take-All)

Instant decisions with <1us latency:

// Enable WTA with 1000 neurons
node.enableWTA(1000);

BTSP (One-Shot Learning)

Immediate pattern association without iterative training:

// Enable BTSP with 128-dim inputs
node.enableBTSP(128);

// One-shot associate a pattern
const pattern = new Float32Array(128);
node.oneShotAssociate(pattern, 1.0);

Morphogenetic Network

Self-organizing network topology through cellular differentiation:

// Enable 100x100 morphogenetic grid
node.enableMorphogenetic(100);

// Network grows automatically
console.log(`Cells: ${node.getMorphogeneticCellCount()}`);

Stepping All Capabilities

In your main loop, step all capabilities forward:

function gameLoop(dt) {
  // Step exotic capabilities
  node.stepCapabilities(dt);

  // Process tasks
  node.processNextTask();
}

setInterval(() => gameLoop(0.016), 16);  // 60 FPS

Building with Exotic Features

# Build with exotic capabilities
wasm-pack build --target web --release --out-dir pkg -- --features exotic

# Build with learning-enhanced capabilities
wasm-pack build --target web --release --out-dir pkg -- --features learning-enhanced

# Build with all exotic capabilities
wasm-pack build --target web --release --out-dir pkg -- --features exotic-full

Core Architecture & Capabilities

Edge-net is a production-grade distributed AI computing platform with ~36,500 lines of Rust code and 177 passing tests.

Unified Attention Architecture

Four attention mechanisms that answer critical questions for distributed AI:

┌─────────────────────────────────────────────────────────────────────────────┐
│                    UNIFIED ATTENTION ARCHITECTURE                            │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐            │
│  │ Neural Attention│  │  DAG Attention  │  │ Graph Attention │            │
│  │                 │  │                 │  │                 │            │
│  │ "What words     │  │ "What steps     │  │ "What relations │            │
│  │  matter?"       │  │  matter?"       │  │  matter?"       │            │
│  │                 │  │                 │  │                 │            │
│  │ • Multi-head    │  │ • Topo-sort     │  │ • GAT-style     │            │
│  │ • Q/K/V project │  │ • Critical path │  │ • Edge features │            │
│  │ • Softmax focus │  │ • Parallelism   │  │ • Message pass  │            │
│  └─────────────────┘  └─────────────────┘  └─────────────────┘            │
│                                                                             │
│  ┌─────────────────────────────────────────────────────────────┐          │
│  │                   State Space Model (SSM)                    │          │
│  │                                                              │          │
│  │     "What history still matters?" - O(n) Mamba-style         │          │
│  │                                                              │          │
│  │  • Selective gating: What to remember vs forget              │          │
│  │  • O(n) complexity: Efficient long-sequence processing       │          │
│  │  • Temporal dynamics: dt, A, B, C, D state transitions       │          │
│  └─────────────────────────────────────────────────────────────┘          │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

Attention Type	Question Answered	Use Case
Neural	What words matter?	Semantic focus, importance weighting
DAG	What steps matter?	Task scheduling, critical path analysis
Graph	What relationships matter?	Network topology, peer connections
State Space	What history matters?	Long-term memory, temporal patterns

AI Intelligence Layer

┌─────────────────────────────────────────────────────────────────────────────┐
│                       AI Intelligence Layer                                  │
├─────────────────────────────────────────────────────────────────────────────┤
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐             │
│  │  HNSW Index     │  │  AdapterPool    │  │  Federated      │             │
│  │  (memory.rs)    │  │   (lora.rs)     │  │ (federated.rs)  │             │
│  │                 │  │                 │  │                 │             │
│  │ • 150x speedup  │  │ • LRU eviction  │  │ • TopK Sparse   │             │
│  │ • O(log N)      │  │ • 16 slots      │  │ • Byzantine tol │             │
│  │ • Cosine dist   │  │ • Task routing  │  │ • Rep-weighted  │             │
│  └─────────────────┘  └─────────────────┘  └─────────────────┘             │
│                                                                             │
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐             │
│  │  DAG Attention  │  │  LoraAdapter    │  │ GradientGossip  │             │
│  │                 │  │                 │  │                 │             │
│  │ • Critical path │  │ • Rank 1-16     │  │ • Error feedback│             │
│  │ • Topo sort     │  │ • SIMD forward  │  │ • Diff privacy  │             │
│  │ • Parallelism   │  │ • 4/8-bit quant │  │ • Gossipsub     │             │
│  └─────────────────┘  └─────────────────┘  └─────────────────┘             │
└─────────────────────────────────────────────────────────────────────────────┘

Swarm Intelligence

Component	Capability	Description
Entropy Consensus	Belief convergence	Shannon entropy-based decision making
Collective Memory	Pattern sharing	Hippocampal-inspired consolidation and replay
Stigmergy	Pheromone trails	Ant colony optimization for task routing
Consensus Coordinator	Multi-topic	Parallel consensus on multiple decisions

Compute Acceleration

┌─────────────────────────────────────────────────────────────────────────────┐
│                      COMPUTE ACCELERATION STACK                              │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│  ┌─────────────────────────────────────────────────────────────────────┐   │
│  │                      WebGPU Compute Backend                          │   │
│  │                                                                      │   │
│  │  • wgpu-based GPU acceleration (10+ TFLOPS target)                   │   │
│  │  • Matrix multiplication pipeline (tiled, cache-friendly)            │   │
│  │  • Attention pipeline (Flash Attention algorithm)                    │   │
│  │  • LoRA forward pipeline (<1ms inference)                            │   │
│  │  • Staging buffer pool (16MB, zero-copy transfers)                   │   │
│  └─────────────────────────────────────────────────────────────────────┘   │
│                                                                             │
│  ┌─────────────────────────────────────────────────────────────────────┐   │
│  │                      WebWorker Pool                                  │   │
│  │                                                                      │   │
│  │  +------------------+                                                │   │
│  │  |   Main Thread    |                                                │   │
│  │  |  (Coordinator)   |                                                │   │
│  │  +--------+---------+                                                │   │
│  │           |                                                          │   │
│  │     +-----+-----+-----+-----+                                        │   │
│  │     |     |     |     |     |                                        │   │
│  │  +--v-+ +-v--+ +--v-+ +--v-+ +--v-+                                  │   │
│  │  | W1 | | W2 | | W3 | | W4 | | Wn |  (up to 16 workers)             │   │
│  │  +----+ +----+ +----+ +----+ +----+                                  │   │
│  │     |     |     |     |     |                                        │   │
│  │     +-----+-----+-----+-----+                                        │   │
│  │           |                                                          │   │
│  │     SharedArrayBuffer (when available, zero-copy)                    │   │
│  └─────────────────────────────────────────────────────────────────────┘   │
│                                                                             │
│  ┌────────────────────────┐  ┌────────────────────────┐                   │
│  │   WASM SIMD (simd128)  │  │   WebGL Compute        │                   │
│  │   • f32x4 vectorized   │  │   • Shader fallback    │                   │
│  │   • 4x parallel ops    │  │   • Universal support  │                   │
│  │   • All modern browsers│  │   • Fragment matmul    │                   │
│  └────────────────────────┘  └────────────────────────┘                   │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

Backend	Availability	Performance	Operations
WebGPU	Chrome 113+, Firefox 120+	10+ TFLOPS	Matmul, Attention, LoRA
WebWorker Pool	All browsers	4-16x CPU cores	Parallel matmul, dot product
WASM SIMD	All modern browsers	4x vectorized	Cosine distance, softmax
WebGL	Universal fallback	Shader compute	Matrix operations
CPU	Always available	Loop-unrolled	All operations

WebGPU Pipelines

Pipeline	Purpose	Performance Target
Matmul	Matrix multiplication (tiled)	10+ TFLOPS
Attention	Flash attention (memory efficient)	2ms for 4K context
LoRA	Low-rank adapter forward pass	<1ms inference

WebWorker Operations

Operation	Description	Parallelization
MatmulPartial	Row-blocked matrix multiply	Rows split across workers
DotProductPartial	Partial vector dot products	Segments split across workers
VectorOp	Element-wise ops (add, mul, relu, sigmoid)	Ranges split across workers
Reduce	Sum, max, min, mean reductions	Hierarchical aggregation

Work Stealing

Workers that finish early can steal tasks from busy workers' queues:

LIFO for local tasks (cache locality)
FIFO for stolen tasks (load balancing)

Economics & Reputation

Feature	Mechanism	Purpose
AMM	Automated Market Maker	Dynamic rUv pricing
Reputation	Stake-weighted scoring	Trust computation
Slashing	Byzantine penalties	Bad actor deterrence
Rewards	Contribution tracking	Fair distribution

Network Learning

Component	Learning Type	Application
RAC	Adversarial Coherence	Conflict resolution
ReasoningBank	Trajectory learning	Strategy optimization
Q-Learning	Reinforcement	Security adaptation
Federated	Distributed training	Model improvement

Self-Learning Hooks & MCP Integration

Edge-net integrates with Claude Code's hooks system for continuous self-learning.

Learning Scenarios Module

use ruvector_edge_net::learning_scenarios::{
    NeuralAttention, DagAttention, GraphAttention, StateSpaceAttention,
    AttentionOrchestrator, ErrorLearningTracker, SequenceTracker,
    get_ruvector_tools, generate_settings_json,
};

// Create unified attention orchestrator
let orchestrator = AttentionOrchestrator::new(
    NeuralAttention::new(128, 4),      // 128 dim, 4 heads
    DagAttention::new(),
    GraphAttention::new(64, 4),         // 64 dim, 4 heads
    StateSpaceAttention::new(256, 0.95), // 256 dim, 0.95 decay
);

// Get comprehensive attention analysis
let analysis = orchestrator.analyze(tokens, &dag, &graph, &history);

Error Pattern Learning

let mut tracker = ErrorLearningTracker::new();

// Record errors for learning
tracker.record_error(ErrorPattern::TypeMismatch, "expected String", "lib.rs", 42);

// Get AI-suggested fixes
let fixes = tracker.get_suggestions("type mismatch");
// ["Use .to_string()", "Use String::from()", ...]

MCP Tool Categories

Category	Tools	Purpose
VectorDb	`vector_search`, `vector_store`, `vector_query`	Semantic similarity
Learning	`learn_pattern`, `train_model`, `get_suggestions`	Pattern recognition
Memory	`remember`, `recall`, `forget`	Vector memory
Swarm	`spawn_agent`, `coordinate`, `route_task`	Multi-agent coordination
Telemetry	`track_event`, `get_stats`, `export_metrics`	Usage analytics
AgentRouting	`suggest_agent`, `record_outcome`, `get_routing_table`	Agent selection

RuVector CLI Commands

# Session management
ruvector hooks session-start    # Start learning session
ruvector hooks session-end      # Save patterns

# Intelligence
ruvector hooks stats            # Show learning stats
ruvector hooks route <task>     # Get agent suggestion
ruvector hooks suggest-context  # Context suggestions

# Memory
ruvector hooks remember <content> -t <type>  # Store memory
ruvector hooks recall <query>                # Semantic search

Claude Code Hook Events

Event	Trigger	Action
`PreToolUse`	Before Edit/Bash	Agent routing, risk analysis
`PostToolUse`	After Edit/Bash	Q-learning update, pattern recording
`SessionStart`	Conversation begins	Load intelligence
`Stop`	Conversation ends	Save learning data
`UserPromptSubmit`	User message	Context suggestions
`PreCompact`	Before compaction	Preserve context

Research Foundation

Edge-net is built on research in:

Distributed Computing - P2P resource sharing
Collective Intelligence - Emergent optimization
Game Theory - Incentive-compatible mechanisms
Adaptive Security - Q-learning threat response
Time Crystals - Floquet engineering for coordination
Neuromorphic Computing - BTSP, HDC, WTA mechanisms
Decentralized Governance - Neural Autonomous Organizations

Disclaimer

Edge-net is a research platform for collective computing. The rUv units are:

Resource participation metrics, not currency
Used for balancing contribution and consumption
Not redeemable for money or goods outside the network

License

MIT License

51 KiB Raw Blame History

@ruvector/edge-net

Table of Contents

What is Edge-Net?

Why Collective AI Computing?

Core Principles

Key Features

Collective Resource Sharing

AI Computing Capabilities

Core AI Tasks

MicroLoRA Adapter System

SONA: Self-Optimizing Neural Architecture

HNSW Vector Index

Federated Learning

ReasoningBank & Learning Intelligence

Pi-Key Identity System

Self-Optimizing Network

Quick Start

1. Add to Your Website

2. Use the Collective's AI Power

3. Monitor Your Contribution

How It Works

The Contribution Cycle

Network Growth Phases

Fair Resource Allocation

AI Computing Tasks

Vector Search (Distributed Similarity)

Embedding Generation

Semantic Matching

Secure Operations

Pi-Key Identity System

Key Types

Using Pi-Keys

Security Architecture

Cryptographic Primitives

Identity Protection

Authority Verification

Attack Resistance

Self-Optimization

Automatic Task Routing

Fitness-Based Evolution

Q-Learning Security

Tutorials

Tutorial 1: Join the Collective

Tutorial 2: Contribute and Earn

Tutorial 3: Use Collective AI Power

Tutorial 4: Monitor Network Health

API Reference

Core Methods

Identity Methods

Network Methods

Development

Build

Test

Run Simulation

Exotic AI Capabilities

Available Feature Flags

Time Crystal (P2P Synchronization)

NAO (Neural Autonomous Organization)

MicroLoRA (Per-Node Self-Learning)

HDC (Hyperdimensional Computing)

WTA (Winner-Take-All)

BTSP (One-Shot Learning)

Morphogenetic Network

Stepping All Capabilities

Building with Exotic Features

Core Architecture & Capabilities

Unified Attention Architecture

AI Intelligence Layer

Swarm Intelligence

Compute Acceleration

WebGPU Pipelines

WebWorker Operations

Work Stealing

Economics & Reputation

Network Learning

Self-Learning Hooks & MCP Integration

Learning Scenarios Module

Error Pattern Learning

MCP Tool Categories

51 KiB

Raw Blame History