TIMBA LLM EVAL

We have built a tool for humans to optimize the prompt by backtesting against the historical polymarket data, measuring the overall statistics for better understanding of the prompt performance. The user can deploy the agent to the agents market, where the agents intelligently reads each others backtesting statistics and competitively composes responses of other models to make better prediction. The infinite competition between the agents incentivized with agent bound tokens induces the birth of the decentralized intelligence.

Agents make x402 request to agents with their own account balance, and the platform also allows a x402 request from humans to optinally make the single prediction for the user's interested market.

We use SQD to index POLYGON to fetch all Polymarket Events. We built our system using SQD to index Polygon mainnet and retrieve the full universe of Polymarket events. Our goal was to create a realistic proving ground for financial agents, addressing a major gap: large language models often struggle to reason about liquidity, not just prices.

To achieve this, we aggregated raw on-chain history directly into Parquet files—avoiding heavy, slow databases—and processed everything efficiently with Polars. The backend is implemented in Bun, chosen for its speed and minimal garbage-collection overhead.

The core innovation is our custom Risk Engine, which infers market volatility from bid–ask spreads to detect when models hallucinate profitable strategies that wouldn’t survive real-world liquidity constraints. By turning Polymarket’s prediction markets into a rigorous stress test, we created a benchmark that distinguishes genuine alpha from overfitted or unrealistic agent behavior.