The discovery layer calls Polymarket's Gamma API with the weather tag filter, turning up anywhere from 200 to 1,200 active books at a given time. Every book holds several temperature outcomes, each with its own market price.

Each title is unpacked into structured metadata — city (mapped to lat/lon), date, temperature unit (°C / °F), and outcome bucket. The parser handles ranges ("30–31°F"), exact values ("14°C"), and tail buckets ("15°C or higher").

Four concurrent API calls to four independent meteorology services, each returning the full hourly temperature array for the target date.

The four forecasts are blended into an ECMWF-weighted consensus with automatic outlier suppression. The probability core then runs horizon-aware sigma, Bayesian historical fold-in, Normal-CDF integration over each outcome bucket, and quarter-Kelly sizing — all to compare our number with the market's implied price.

Claude is prompted as a senior meteorologist with 15 years on the desk. Every analysis runs a five-step protocol: individual model review with known biases, synoptic pattern read, microclimate adjustments (urban heat island, coastal effects), Weather Underground station awareness, and a side-by-side comparison against current market pricing. Running on Claude Sonnet 4.6.

Orders route through Polymarket's CLOB as Fill-Or-Kill, falling back to GTC limits if needed. The streaming layer watches every open position in real time; the exit monitor enforces stop-losses, trailing stops and profit targets. Positions and balances reconcile with Polymarket every two minutes.