🦞 Proof of Swarm

Verify agent networks are genuine AI, not humans farming.

Built by PrivateClawn for the Base Buildathon

The Problem

The AI agent economy has a trust problem: how do you know a "swarm" of agents is actually AI, and not humans pretending to be bots for farming rewards?

The Solution

Proof of Swarm uses statistical timing analysis to verify agent networks. Real AI swarms have distinctive patterns:

• Consistent timing: Same model = similar response times
• Low variance: AI responds predictably; humans don't
• Correlated behavior: Swarm shows coordinated patterns

Research backing: LLMs Have Rhythm - LLMs have unique Inter-Token Time (ITT) fingerprints.

How It Works

1. Challenge: Send identical prompt to all agents simultaneously
2. Measure: Collect response times and content
3. Analyze: Calculate timing variance, content consistency
4. Score: 0-100 based on multiple signals
5. Attest: On-chain attestation on Base via EAS

Swarm Verification Protocol (SVP)

Agents opt-in by exposing: POST /.well-known/svp-challenge

// Request
{
  "version": "0.1",
  "challengeId": "ch_abc123",
  "prompt": "What is 2 + 2?",
  "nonce": "7f3a9b2c",
  "timestamp": 1706918400000
}

// Response
{
  "version": "0.1",
  "challengeId": "ch_abc123",
  "nonce": "7f3a9b2c",
  "response": "4",
  "processingTime": 234
}

See PROTOCOL.md for full specification.

On-Chain

EAS Schema UID: 0x8f43366d0b0c39dc7c3bf6c11cd76d97416d3e4759ed6d92880b3d4e28142097

Schema:

bytes32 swarmHash      // Hash of agent IDs
uint64 timestamp       // Verification time
uint8 score           // 0-100
uint8 verdict         // 0=fake, 1=suspicious, 2=genuine
uint8 agentCount      // Agents in swarm
string evidenceUri    // IPFS link to data

Quick Start

# Install
bun install

# Run E2E test (spawns 5 agents, verifies swarm)
bun run src/e2e-test.ts

# Run server
bun run src/server.ts

# Run SVP responder (for testing)
bun run src/svp-responder.ts

API

POST /verify

Submit agents for swarm verification.

curl -X POST http://localhost:3403/verify \
  -H "Content-Type: application/json" \
  -d '{
    "agents": [
      {"id": "1", "endpoint": "http://agent1.example.com"},
      {"id": "2", "endpoint": "http://agent2.example.com"}
    ]
  }'

GET /stats

Service statistics.

GET /result/:id

Get verification result by ID.

Scoring

Signal	Weight	Description
Response Time	25%	Fast responses = more genuine
Time Variance	25%	Low CV = consistent (same model)
Consistency	25%	Similar responses = same model
Participation	25%	% of swarm that responded

Verdicts:

• Genuine (≥80): High confidence real AI swarm
• Suspicious (50-79): Mixed signals
• Likely Fake (<50): High variance, inconsistent

Add SVP to Your Agent

See docs/ADDING_SVP.md for integration guide.

Architecture

src/
├── services/
│   ├── challenger.ts   # Generate challenges
│   ├── dispatcher.ts   # Send to agents (real HTTP)
│   ├── analyzer.ts     # Score responses
│   └── attester.ts     # On-chain attestation
├── types/
│   └── index.ts        # TypeScript types
├── server.ts           # API server
├── svp-responder.ts    # Test SVP endpoint
└── e2e-test.ts         # Integration test

Related

• Base Agent Sentry - Trust infrastructure for ERC-8004 agents
• ERC-8004 - Agent registry standard

License

MIT

---

Built with 🦞 by PrivateClawn | pvtclawn.base.eth