Farcaster native game (mini app) where players wager ETH spotting AI clones. Built on Arbitrum with Solidity + Rust for on-chain reputation & Proof of Humanity.
Detective is a Farcaster-native social deduction game tackling one of the internet’s most urgent challenges: proving humanity in an age of AI. As models grow indistinguishable from real users, “Proof of Humanity” becomes a scarce and valuable primitive. Detective transforms this problem into a competitive, onchain experience where players wager ETH or USDC to decide whether they’re interacting with a real person or an AI agent trained on that user’s Farcaster history.
Each match produces high-signal human feedback—crowdsourcing adversarial data to combat synthetic identity at scale. We turn AI detection into a game: playable, measurable, and economically incentivized.
Detective is architected natively on Arbitrum with a hybrid smart contract design optimized for performance and cost-efficiency:
Solidity (economic layer): Secure entry fees, staking, and trustless pull-payment withdrawals.
Arbitrum Stylus (Rust): High-compute reputation logic including Deception Success Rates and dynamic Humanity Scores.
By offloading complex adversarial metrics to Stylus, we achieve order-of-magnitude efficiency gains without inflating gas costs. Our live deployment on Arbitrum Sepolia demonstrates a production-ready system handling staking, settlement, and Sybil-resistant verification fully onchain.
Detective evolves beyond a game into infrastructure:
Phase 1 (Live): PvP human-vs-bot matches with real-time chat and voting.
Phase 2: Public Agent Leaderboard ranking AI clones by deception performance.
Phase 3: Protocol API enabling any Arbitrum dApp to query onchain Humanity Scores for wallet-level verification.
We’re building the Turing Oracle for Arbitrum’s agent economy—a decentralized intelligence layer that makes identity verifiable, reputation programmable, and AI detection economically aligned.
<p>During the hackathon, we moved Detective from concept to a production-ready, Arbitrum-native protocol with live contracts, Stylus integration, and AI fine-tuning.</p><p>🦀 1. Arbitrum Stylus: High-Compute Reputation in Rust</p><p>We implemented a hybrid architecture:</p><p>Solidity (Arbitrum One): Handles entry fees, staking, settlement, and Sybil-resistant registration.</p><p>Stylus (Rust/WASM): Powers high-compute adversarial metrics including:</p><p>Deception Success Rate (DSR)</p><p>Dynamic Humanity Scores</p><p>Cross-round behavioral analysis</p><p>By moving complex scoring logic into Stylus, we achieved significantly more efficient computation for adversarial reputation models—without inflating user gas costs. This enables scalable, onchain intelligence rather than offchain black-box scoring.</p><p>🔐 2. Live Arbitrum One Deployment</p><p>We deployed and verified our production contract on Arbitrum One, implementing:</p><p>Trustless pull-payment withdrawals (V4 architecture)</p><p>One-wallet-per-cycle Sybil resistance</p><p>Onchain event tracking for traction metrics</p><p>Admin pause controls and configurable entry fees</p><p>This isn’t a mock deployment—players must sign a real Arbitrum transaction to enter the arena.</p><p>🤖 3. AI Fine-Tuning & Identity Cloning</p><p>We built a full AI identity-cloning pipeline:</p><p>Scrape 30+ recent Farcaster casts per user (via Neynar)</p><p>Extract 20+ personality traits (tone, cadence, emoji patterns, topics)</p><p>Inject structured behavioral priors into Claude 3.5 Sonnet</p><p>Enforce Farcaster-native constraints (≤240 chars, conversational rhythm)</p><p>Enhancements shipped during hackathon:</p><p>Realistic 2–7s typing delays</p><p>Personality-weighted opening moves</p><p>Authentic fallback generation strictly from cast history</p><p>Cross-round memory using Redis-backed lightweight context</p><p>Multi-model experimentation (Claude + Llama 3.3)</p><p>Result: Bots that genuinely feel like the user they’re cloned from—raising the difficulty and improving the quality of adversarial training data.</p><p>🔗 4. Farcaster Native Integration (2025 Standard)</p><p>We migrated fully to Farcaster Quick Auth:</p><p>Edge-signed JWT verification (no nonce juggling)</p><p>Auto-approval inside Warpcast</p><p>73% build size reduction after removing legacy wallet dependencies</p><p>Mini App SDK integration with proper ready signals</p><p>Detective runs as a true Farcaster-native Mini App—not a web app wrapped in crypto.</p><p>⚡ 5. Real-Time Multiplayer Infrastructure</p><p>We upgraded the gameplay stack with:</p><p>WebSocket implementation (Ably) with feature flags</p><p>Registration lobby + countdown ceremony</p><p>Multi-chain leaderboard architecture (Arbitrum + Monad prep)</p><p>Modular access gating for NFT/token-based entry</p><p>The system is horizontally scalable and structured for future Agent Economy expansion.</p><p>📊 6. Measurable Onchain Traction</p><p>Entry transactions recorded on Arbitrum</p><p>Smart contract events used for analytics</p><p>Leaderboard rankings computed from verified match outcomes</p><p>Fully passing TypeScript strict build (Next.js 15)</p><p>This is not a demo-only UI—it’s a functioning onchain game generating adversarial AI detection data today.</p><p>🧠 What We Proved During the Hackathon</p><p>Stylus can power compute-heavy reputation logic efficiently.</p><p>Onchain identity primitives can be gamified.</p><p>AI detection can be economically incentivized.</p><p>Arbitrum can host the “Turing Oracle” layer for the agent economy.</p><p>Detective evolved from a game concept into infrastructure: a decentralized intelligence layer where humanity is measurable, reputation is programmable, and AI deception becomes economically visible.</p>
<p>n/a</p>
