LosCommosHermanos

Swath nowcasts corn and soybean yields for every Iowa county before the harvest, from public data only, with an honest interval around each figure. Satellite vegetation and weather signals feed a trend-plus-anomaly model; each county figure comes with a calibrated 80% interval and attributions that say which signals moved it, rolls up to the state, and is set against USDA's monthly forecast. Meta's models take it from pixels to fields: SAM 3.1, on Meta's Model API, is outlining Iowa's fields on public aerial imagery, SAM 3 does the same in our scheduled pipeline, DINOv3 embeddings label each field corn, soybeans or other (preliminary), and Muse Glimmer checks the least certain outlines. Field signals enter the nowcast only once our point-in-time benchmark accepts them; it keeps a public record of what it rejected. Over 2014–2025, the September state error is 2.7% against USDA's 3.7%: comparable, not significantly better. An analyst agent, a local open model, answers from these outputs only. Most of the public code was built during the event. Live, no sign-in: swath.qbuissondebon.eu.org DISCLAIMER REGARDING THE TEAM COMPOSITION I want to make it clear that "c0mmod1t1" has worked in collaboration with "automation_lord" on this project. "automation_lord" could not be added to the team as he appears not to be an approved attendee. He could no longer register to the event today, but agrees and accepts all the Official Rules of the Event. I have tried contacting the organisers but could not have an answer in time to help solve this issue: - https://discord.com/channels/1557169028080869426/1567658844253978756 - https://discord.com/channels/1557169028080869426/1567664163168321566 We apologise for the confusion. We hope this will not pose a risk of disqualification, as we had decided to do this challenge together from the get-go, we both acted in good faith throughout the duration of the Event, and may have committed an honest mistake during the registration process as we're just realising now.