A looped system that uses an Anthropic-based LLM research agent to sweep the internet for complete, up-to-date datasets on each prediction market. That data together with a quantitative prediction model is fed to JEV, the System One probability engine, which outputs a calibrated probability for every market.
When JEV's probability is an outlier against the crowd's price, the system places a fixed-stake bet on the market from a Solana-based wallet with a $2,500 starting balance. Open positions are reviewed roughly every 6 hours, to either close early, or wait for resolution.
Every judgment is recorded to review whether the machine actually beats the masses, and/or to decide whether or not to replace/improve the research agent.
the research agent · Anthropic …
When a market enters the queue, Scout sweeps the internet including official data sources, news and filings, and assembles a frozen, hash-verified evidence package: facts, base rates, external forecasts and directional signals, every claim with a source. On first contact with a new market type, Scout also writes the reusable research template for that type.
Scout never sets the probability. It only gathers what JEV needs.
status: checking…
the decider · TypeSafe System One
JEV is a calibrated decision model, not a chat LLM. It reads the frozen evidence package and answers exactly one question per market: the probability it resolves YES. A single scoreable number, no browsing, no prose.
When JEV's number is an outlier against the crowd's price, the arena bets a fixed stake. Every judgment lands on the scoreboard, win or lose.
status: checking… · model …
| market | side | entry | stake | bid now | unrealized | settles |
|---|---|---|---|---|---|---|
| loading… | ||||||
| model | n | brier | log loss |
|---|---|---|---|
| loading… | |||
| 1 Research | live datasets per market |
| 2 Judge | JEV System One → probability |
| 3 Bet outliers | fixed stake vs the crowd |
| 4 Review | every few hours: close or hold |
| 5 Score | every outcome, win or lose |