Prediction-market resolver.

Existing prediction-market oracles have three structural flaws: resolution becomes truth by inertia if no one challenges a stale proposal within the short window; challenging costs a bond that prices out small holders; and final disputes are decided by token weight, not evidence.

Our team shipped an alternative: an AI agent bound by the market's own rules (injected as system-prompt context), given real-time web search to gather authoritative sources, and forced to call a `submit_resolution` tool with outcome + reasoning + citation list + confidence. Every decision produces a full audit memo the platform can publish.

Each cycle begins with the resolver pulling the market's question, rules, allowed outcomes, and primary sources. A dynamic prompt is generated from that context and pushed into the system prompt. The AI runs the web-search plugin, evaluates the evidence against the resolution criteria, and submits the structured outcome — no string parsing on the calling code's side, the tool call *is* the contract.

Confidence below a configured threshold auto-resolves to UNKNOWN rather than guessing. Outcomes are stored with processing time, model used, and a full reasoning trail so any dispute can be replayed. Production-ready upgrades in the roadmap message queue replacement (Kafka / RabbitMQ), multi-model consensus, persistent PostgreSQL backing the resolution store, and rate limiting on the resolve endpoint.