Hard, API-verified findings from a deep 5-agent technical audit of weather-bet on 2026-06-07, checked against live Polymarket APIs, vendor docs, and the polymarket-fetch source code. Several of these correct earlier assumptions in polymarket-weather-market-structure and weather-bet — treat this note as the authoritative override where they conflict.
For Agents
Everything below was verified live on 2026-06-07: against the Gamma/CLOB APIs, vendor documentation, or by reading
polymarket_fetchcode on disk. Where this note contradicts the older market-structure note, this note wins (it post-dates and corrects it). Items marked CORRECTION overturn a prior belief. Items in[!danger]/[!warning]callouts are silent-failure or blocker risks — read those before writing any resolution, calibration, or execution code.
Two findings can silently lose money — read these first
- Resolution rounding (see Resolution ground truth (CRITICAL — silent-failure)): computing daily max from raw METAR tenths-°C disagrees with Polymarket’s actual whole-°F-derived resolution ~16% of the time, exactly at half-degree bucket boundaries. You must reconstruct ground truth from whole-°F observations.
- Ensemble underdispersion (see Forecasting science risks): raw NWP ensembles are 3–4× underdispersed at day-1. Mapping raw members straight to buckets is catastrophically overconfident. Calibration (EMOS/NGR) is mandatory, not optional.