This is the 20260630 visual detection release, covering both the video and image models.The video model backbone moved to V-JEPA 2. F1 went from 0.985 to 0.990 and AUC from 0.993 to 0.997. On the image side, detection of passport and document fraud improved, with average accuracy up from 84.4% to 92.9%.The report itself was refreshed too. The heatmap now uses edge-highlighting, and the timeline tree-view adds representative-frame analysis, so it is easier to see where and why the model flagged something.