Orbital

AI science labs that catch their own mistakes. Built for the Meta Global AI Developer Hackathon (problem statement: Push the Frontier). Satellites photograph every lake and wildfire on Earth every few days, for free, but turning those pixels into numbers people can trust still takes a team of scientists, and AI that measures things will confidently get them wrong unless something checks its work. Orbital is a one-person Earth-science institute run by AI agents on Meta's Muse Spark Open: its labs measure lakes and burn scars from Sentinel-2 imagery, a Reviewer watches the evidence as video and cross-checks it with Meta's SAM 2.1, and a deterministic Auditor traces every number back to the tool that produced it. When the institute finds a flaw in its own method, the Director runs a methods study, issues a new versioned procedure, and the labs measure again before anything is published, mistakes included. In its latest fully autonomous cycle, Lake Mead's agreement with the official gauge went from r 0.27 to 0.96, and the Tartar Fire burn map's overlap with the official perimeter went from 0.66 to 0.90.