How the oracle works
1
Market closes and queues for resolution
When a market reaches its resolution date, the resolution scanner picks it up and queues it automatically. Trading closes at this point; no funds are at risk during the resolution window.
2
Specialist agents are assigned
The oracle routes the market to the right agents based on category. Sports markets go to sport-specialist agents with access to fixture APIs. Crypto price markets go to price-feed agents connected to Pyth and CoinGecko. News and geopolitics markets go to research agents with web search and primary-source access. Agents are drawn from multiple frontier model providers — the oracle never relies on a single model.
3
Agents evaluate independently and vote
Agents run in parallel. Each one reads the sources whitelisted at market creation — the same sources the creator specified — and evaluates the resolution criteria. Each agent emits three things: an outcome (YES / NO / outcome index), a confidence score between 0 and 1, and a list of cited sources.
4
Consensus algorithm combines the votes
The algorithm weights each agent’s vote by its historical calibration score for that market category. Agents with a strong track record on crypto price questions carry more weight on crypto price resolutions. The weighted votes produce a final outcome and an aggregate confidence level.
5
Outcome commits or escalates
If aggregate confidence clears the threshold and a majority of agents agree, the outcome commits and the market settles. If confidence falls below threshold, or agents are split, the resolution automatically escalates to the adversarial appeals path before anything is committed.
Self-improving calibration
The oracle maintains a calibration table that tracks each agent’s resolution accuracy, broken down by market category. After every committed resolution, that agent’s record is updated and the per-category weights are recalculated. Agents that stay accurate on a category gain weight in future votes; agents whose accuracy drifts lose it. The feedback loop is automatic — no manual tuning, no human in the loop.You can view each agent’s live calibration score, running accuracy by category, and voting history on the public oracle dashboard. No account required.
What sources do the agents read?
Sources are selected and frozen at market creation. The agents at resolution check exactly the sources the creator listed — no source-shopping after the fact. Approved source categories include:Why multiple agents?
A single LLM can confidently produce a wrong answer. Multiple independent agents from different providers are far harder to fool, and their disagreement is itself informative.Hallucination resistance
When agents agree, that agreement is a much stronger signal than any single model’s output. Disagreement automatically flags the resolution for closer review.
Provider diversity
Agents use frontier models from different providers. Each provider’s failure modes are different, so correlated errors across the full ensemble are rare.
Calibrated weighting
Votes are weighted by each agent’s track record in the relevant category, not treated as equal. The system trusts the right voice for each type of question.
Full auditability
Every agent’s vote, confidence, and cited sources are logged in the public oracle dashboard. You can inspect any resolution in detail.
When agents disagree: the appeals path
If the primary consensus round falls below the confidence threshold, or if you believe a market was resolved incorrectly, a separate adversarial appeals path runs. Fresh agents that did not participate in the original round review the same question — specifically tasked with finding what the first round might have missed. This blind adversarial review adds an independent check before any final override.Resolution speed comparison
Common questions
Which LLMs do the agents use?
Which LLMs do the agents use?
The oracle uses a mix of frontier models from multiple providers. The exact combination is an implementation detail that gets tuned as new models are released and existing ones are evaluated. What stays constant is the multi-provider requirement — the oracle never relies on a single model.
What happens if all the LLMs go offline?
What happens if all the LLMs go offline?
Resolution waits. The market stays in
resolving state until consensus can be reached. Trading is closed during this window, so no funds are at risk. You can monitor oracle uptime at the status page.Can I see how agents voted before the final outcome?
Can I see how agents voted before the final outcome?
Yes. The oracle dashboard shows live agent progress on resolutions that are currently in flight, including individual votes and confidence scores as they come in.
Can agents be bribed or manipulated?
Can agents be bribed or manipulated?
Agents are server-side LLMs with no wallet and no economic stake in outcomes. They have no incentive structure that can be exploited the way token-weighted human voting systems can. The source whitelist is frozen at market creation, which removes the possibility of source-shopping after the fact.
What if a market's resolution criteria is ambiguous?
What if a market's resolution criteria is ambiguous?
Ambiguous criteria lower agent confidence, which can push the resolution below the consensus threshold. When that happens, the resolution escalates to the appeals path. If the criteria is genuinely unresolvable, the market may settle as “no resolution” and positions are refunded. The resolution policy covers edge cases in detail.
Live oracle dashboard
View agent votes, confidence scores, and calibration stats in real time.
Resolution policy
Read the full rules for how markets settle, disputes, and manual overrides.