V1.0 - GENESIS

Experiment #1: Can pure reasoning beat prediction markets? We deployed 2 adaptive agents into Polymarket with zero special training, identical prompts, and real capital to find out.

The Hypothesis: Traditional benchmarks test memorization. Prediction markets test understanding. If frontier models can synthesize information, assess risk, and reason strategically under real uncertainty, they should find profitable mispricings.

What We're Testing: Do LLMs have genuine market intuition, or do they just sound confident? Can they adapt their strategies based on past outcomes? Will different architectures develop distinct trading personalities? Real money = real answers.

Current Status: GPT-5 and Claude Sonnet 4.5 actively trading (2/6 deployed). Four additional models (Gemini 2.5 Pro, Grok 4, DeepSeek V3.1, Qwen3-Max) configured and ready for Phase 2 deployment.

The AI agents operate with complete autonomy. No human intervention. No forced check-ins. Each model decides when to check markets again, from 5 minutes to 24 hours.

Every decision includes full context: current wallet balance (queried from blockchain), all open positions with live P&L, last 5 decisions with outcomes, and fresh market data. The AI learns from its history and evolves its strategy autonomously.

Agents don't just trade on instinct. They research markets using advanced web search capabilities via Tavily API.

Before making a trade, agents can perform up to 10 web searches, gathering news, expert opinions, historical data, and market context. Multi-turn research loops allow iterative investigation, enabling deep analysis before committing capital.

Each decision is powered by a sophisticated full-context prompt system that provides the AI with everything it needs to make informed trades.

Every decision includes a confidence score (0.0-1.0), detailed reasoning, and specific action parameters. The system logs every prompt, every response, token usage, and tool calls for complete auditability.

Polymarket has hundreds of prediction markets, but not all are worth trading. We filter 150+ markets down to 20-30 quality options each time.

Prices update every 3 minutes, so models always see current market conditions and know exactly how their positions are performing in real-time.

When a model decides to trade, it happens for real on the blockchain. Real cryptocurrency. Real prediction markets. Real profits and losses.

No simulations or fake money here. Each model has its own blockchain wallet, and every time we check their balance, we query the blockchain directly—the ultimate source of truth. What you see on the dashboard is what's actually on-chain.

Every open position is monitored constantly. Prices refresh every 3 minutes, showing real-time profit or loss on each bet.

For each position, we track the entry price (what they paid), current market price, how many shares they own, and their gain or loss both in dollars and percentage. The AI sees these numbers before every decision—helping it learn which bets worked and which didn't.

Nothing happens in secret. Every decision, every trade, every piece of reasoning—all saved permanently.

This complete record lets us understand how each AI thinks. Why did GPT-5 bet on one market while Claude avoided it? What research did they do? How confident were they? Having this data means we can learn what strategies actually work in real-world trading.

The system runs in three places that work together: a trading computer, a cloud database, and this website.

Trading happens fast on a dedicated computer with its own database. After each trade, key data gets copied to a cloud database that this website reads from. The sync happens in the background, so it never slows down trading.

To prevent disasters, we have built-in limits on what the AI can do. They can still lose money, but they can't blow up the entire portfolio on one bad bet.

No single trade can risk more than 15% of their money—$30 max on a $200 starting balance. And they can't place a bet unless they're at least 60% confident. These rules keep the experiment meaningful without letting things spiral out of control.