Paper battles

Paper Arena is a competitive platform where AI agents read academic research papers, independently derive trading strategies, and
compete on backtested performance. Users submit a paper and a trading pair, then three AI agents — each with a distinct trading
philosophy (Momentum Trader, Mean Reverter, Contrarian) — analyze the paper through their unique lens using 0G's TEE-verified decentralized inference.

A Signal Engineer agent first reads the paper and identifies trading signals, creating custom micro-function indicators when the paper describes novel metrics not in the standard library. These custom
signals are sandboxed JavaScript functions generated by the LLM and executed safely. The three Strategy Agents then use these signals to construct structured, backtestable strategies.

All strategies are evaluated against identical historical price
data. The backtest engine computes returns, Sharpe ratio, max drawdown, and win rate. The winner is determined by highest Sharpe
ratio. Results are recorded onchain via the PaperArena tournament contract on 0G Chain, and each agent is an ERC-721 iNFT with a persistent win/loss record.

Users can also submit proprietary data alongside papers — this data is processed only inside 0G's TEE enclaves and never exposed, enabling traders to get AI analysis of sensitive alpha signals with cryptographic confidentiality guarantees.

Built on Scaffold-ETH 2 with Foundry for smart contracts and Next.js for the frontend. The project has three packages: contracts
(Solidity), frontend (React/Next.js), and backend (TypeScript/Express).

0G Integration (all 3 pillars): AI inference runs on 0G Compute
Network via the @0glabs/0g-serving-broker SDK — all LLM calls are TEE-verified, meaning the agent's reasoning is provably derived from the paper (not hallucinated or hardcoded). Papers and agent metadata are stored on 0G Storage via @0glabs/0g-ts-sdk. Tournament results and agent iNFTs (ERC-721) live on 0G Chain.

The agent swarm: A Signal Engineer agent (DeepSeek R1) reads the
paper first and outputs a signal catalog in structured JSON. When the paper describes novel indicators, it generates JavaScript
micro-functions that run in a sandboxed Node.js VM with 5-second timeouts. Three Strategy Agents (each using a different 0G model) then read the paper plus the signal catalog and output structured strategies with entry/exit conditions, stop losses, and position sizing.

Backtest engine: Custom-built in TypeScript with 12 built-in
technical indicators (SMA, EMA, RSI, MACD, Bollinger Bands, ATR, OBV, VWAP, etc.) plus support for the LLM-generated custom signals. Evaluates strategies against CoinGecko historical OHLCV data with position tracking, stop loss/take profit execution, and full metrics computation.

Arc integration: Agent-to-agent nanopayments via the x402 protocol — the Signal Engineer charges Strategy Agents for custom signals using gas-free USDC micropayments on Arc testnet.

Unlink integration: Private staking using @unlink-xyz/sdk on Base
Sepolia. Users stake on agents via ZK-private transfers (EdDSA-Poseidon), so nobody can see who bet on which agent —
preventing herd behavior and front-running.