Kairosity Prophet Arena Agent was inspired by the core challenge of Prophet Hacks: forecasting systems are only useful when they are both thoughtful and operationally reliable. The endpoint therefore optimizes for two things at once: a disciplined forecast workflow and a strict API contract that always returns valid probabilities.
The agent uses a compact stepwise forecasting rubric:
- Read the resolver carefully.
- Preserve the exact event outcome labels.
- Anchor on current state and relevant priors.
- Build a base rate for the remaining time window.
- Apply evidence updates and calibration.
- Return machine-readable probabilities.
The calibration rule for binary questions is intentionally simple:
$$ X = \min(10 - c, 0.2p) $$
where (p) is the raw percentage estimate and (c) is confidence on a 0-10 scale. The agent applies this idea carefully: it avoids mechanical over-adjustment, and for categorical questions it instead returns a full normalized probability vector over every exact label.
The biggest lesson was that forecasting quality is not just model quality. The surrounding contract matters: exact labels, sane fallbacks, timeout safety, and provider failure handling are part of the forecasting system. A beautiful forecast that misses the schema is still a failed forecast.
The main challenge was keeping the public submission small while preserving the most useful parts of a larger agentic forecasting architecture. The final repo intentionally omits private analysis artifacts and keeps only the deployable endpoint, prompt, tests, and operational instructions.
Log in or sign up for Devpost to join the conversation.