Full guide: Jev 1.14: what changed, when jev-latest moves, and the danger zone
A Box is a named question with its type set to box. It takes instructions naming the target and the image, such as the coffee mug in `photo` , and accepts no criteria. The docs’ example answer is {"type":"box","box":{"top":0.25,"bottom":0.82,"left":0.44,"right":0.84},"confidence":0.92}. Edges are measured from the top-left corner, so the same answer works at any image size. Multiply left and right by the width and top and bottom by the height to get pixels.
Only jev-1.14.0, released 14 October 2026, answers Box questions. Call it by name or through the jev-preview alias, and send the TypeSafe-Danger-Zone-Images: true header with every request that carries an image or a Box question. TypeSafe labels images and Box “Dreamer-ready”, meaning not ready for production: the request format, answers and limits can change between releases.
Box always returns a box, even when the target is not in the picture. Pair each one with a Noul about the same target and ignore the box when the Noul comes back low. Each Box finds one target, so ask separate questions for separate things, and describe the one you want when several look alike.
The docs set no confidence threshold. Below whatever threshold you pick, they suggest asking again on a larger or cropped image, padding the box before you crop, or cropping to the box and asking one Noul per edge about whether the target is cut off. On TypeSafe’s test set of 36 painting crops, those four edge questions caught 30 cut-offs, against 26 for one “is all of the target inside” question.
For a small target in a large image, find the region first, for example with a Choice over image tiles, then run Box on that tile and convert the edges back to full-image fractions. Request examples, image formats and size limits are in the Jev 1.14 guide.