AI-Resolved Prediction Market
Polymarket uses UMA's optimistic oracle — a human-driven dispute mechanism that takes hours and can swing on token-weighted voting. Kalshi has internal staff resolution. BlockForecast does it differently: multi-agent AI consensus from real LLMs reading real public sources.
The architecture (plain English)
- Trigger. Market reaches its resolution date. The resolution scanner queues it.
- Batch parallel. Seven independent agents run at once. Each agent has its own LLM (frontier models from different providers) plus tool access — sports APIs for fixtures, Pyth and CoinGecko for crypto prices, web research for news.
- Specialist routing. Sports markets get sport-specialist agents. Crypto markets get price-feed agents. News markets get research-tool agents. The right agents look at the right kind of question.
- Confidence-weighted vote. Each agent emits an outcome (YES / NO / outcome-index) plus a confidence (0–1) plus cited sources. The consensus algorithm weights each agent's vote by its historical calibration on that category.
- Self-improving. A calibration table tracks each agent's accuracy per category over time. Weights update from realized resolutions. Agents that are accurate gain weight; agents that drift lose it.
- Health monitoring. Agreement rates, confidence drift, source reachability — all tracked continuously. A failing agent auto-suspends without admin intervention.
- Disputes. If consensus is below threshold, or any user files an appeal, a separate blind adversarial appeals path runs — fresh agents that don't see the original consensus. Their job is specifically to look for what the first round missed.
Why multi-agent consensus
- Hallucination resistance. A single LLM can confidently get something wrong. Seven agents that disagree is a flag; seven that agree is a strong signal.
- Provider diversity. Frontier models from different providers fail in different ways. Mixing them collapses the correlated-failure risk.
- Calibration. Tracking each agent's track record per category lets the system trust the right voice for the right kind of question.
- Auditability. Every agent's vote, confidence, and citations are logged on a public dashboard at /oracle.
What the agents read
- Sports: API-Football for fixture results, ESPN, league official sites, Kalshi orderbook (cross-check)
- Crypto: Pyth Hermes (real-time price feed), CoinGecko, Coinbase API
- News / politics: Web search, AP News, Reuters, official sources cited at market creation
- Tech / culture: Web search, official announcements, primary sources
Sources are selected at market creation and frozen — the agents at resolution check the same sources the creator promised. No source-shopping.
Resolution speed
| Mechanism | Typical resolution time |
|---|---|
| UMA optimistic oracle (Polymarket) | ~2 hours minimum (challenge window) |
| Human staff (Kalshi) | Minutes to hours, depends on category |
| Multi-agent AI (BlockForecast) | Typically under 5 minutes |
What if the AI is wrong?
Three layers of safety:
- Threshold. Resolution only commits if multiple agents agree above a confidence threshold. Below threshold → automatic appeals path.
- User appeals. Any user can file an appeal during the appeal window. The appeals path uses fresh agents that don't see the original outcome.
- Manual override. If the AI consensus is provably wrong, the platform reserves manual override as a final escape valve. Used rarely; logged publicly. Full resolution policy.
Transparency
Every resolution is auditable at /oracle. You can see:
- Which agents voted, their individual votes, their confidence scores
- Sources each agent cited
- The consensus algorithm output
- Each agent's running calibration score for the category
- The full audit log
FAQ
Which LLMs do the agents use?
A mix of frontier models from different providers. The exact mix is an implementation detail that gets tuned over time as new models come out and old ones drift.
What if all the LLMs go offline?
Resolution waits. The market stays in resolving state until consensus is reached. Trading is closed during this period; no funds at risk.
Can I see how an agent voted before final consensus?
Yes — the oracle dashboard shows live agent progress on resolutions in flight.
Are agents bribable?
The agents are server-side LLMs running on our infra; they have no wallet and no economic stake. They aren't bribable in the way human voters in token-voting systems are.