Time-Fold

Time-Fold turns two satellite images of the same place, from different years, into evidence about what changed, and checks how far that evidence can be trusted. Meta's SAM 3.1 detects and outlines objects such as buildings, and I turn its masks into counts and area shares. Muse Spark then audits those results: it looks at the original images and the detection overlays, says which detections look right or wrong, flags large structures or whole classes the detector missed, and writes a short report with causes labelled as hypotheses, a qualitative 2050 scenario, and confidence and limits. Tested on Kempegowda International Airport (Bengaluru, 2015 vs 2026) and Tesla Gigafactory Texas (Austin, 2015 vs 2025). In both, the audit caught real detector mistakes. It helps people read satellite change without expertise, and shows which automatic numbers to ignore. Shorter version, if the box has a small limit: Time-Fold compares two satellite images of the same place from different years. Meta's SAM 3.1 detects and measures objects such as buildings, then Muse Spark audits those measurements, flags what the detector got wrong or missed, and writes an honest report with causes as hypotheses and clear limits. Tested on Bengaluru's airport and Tesla Gigafactory Texas.