Docs
OfficialJev Practices & Patterns
Box Questions (Dreamer-Ready)
The box question type: one bounding box per question, returned as four edges that are fractions of the image, plus confidence, with no criteria field and no not-found value.
- Format
- Docs
- Reading time
- 3 min
- Published by
- TypeSafe AI
- Last verified
- Oct 2026
Highlights
- A Box answer is four edges — top, bottom, left, right — each 0 to 1 as fractions of the image, plus a confidence; the same answer works at any image size.
- Box takes no criteria: instructions describe what to find and in which image, and may be a phrase like the coffee mug in photo.
- Box needs jev-1.14.0 or jev-preview plus the images header; jev-1.13.0 returns HTTP 400 for a Box question.
- There is no not-found value: an absent target still returns four edges, so the docs recommend pairing Box with a Noul that checks whether the target is present.
- Reading a Box is arithmetic: multiply each edge by the image width or height and crop, pad, or re-ask at higher resolution when confidence is low.
Watch out
Dreamer-ready: the page says the request format, answers, and limits can change between releases and accuracy changes as models update. There is no recommended confidence threshold yet, so tune one on your own images, and low-confidence answers need a re-ask or padding strategy.
Questions about this entry
Short answers drawn from the linked project or page.
One bounding box per question: box with top, bottom, left, and right, each a fraction of the image from the top-left corner, plus a confidence from 0 to 1. The response also names the model and token usage.
Jev 1.14: send jev-preview or jev-1.14.0. Box is not available on jev-1.13.0, and every request containing a Box question also needs the TypeSafe-Danger-Zone-Images: true header.
Ask a separate Noul question that checks whether the target is present. Box has no not-found value, so an absent object still returns four edges; the docs pair the two so code can reject a box for a missing target.
Multiply each edge by the image's width or height in pixels: left and right by width, top and bottom by height, then crop with those rounded coordinates. The docs show a PIL example cropping a mug out of a kitchen photo.
There is no recommended threshold yet, so tune one on your own images. The docs suggest re-asking on a larger image or a higher-resolution crop, padding the box outward before cropping, or checking each edge with follow-up questions.
Yes. A request can mix Box with Choice, Score, and Noul, and every question is evaluated against the same state. One common pattern is a Box plus a Noul presence check plus per-edge questions for cut-off detection.
Yes. The header is required on every request that contains an image or a Box question, regardless of how the image was sent, whether as a multipart file part or base64 in the JSON body.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
X: First among models under 10B on the Decision Index v0.2.1. Liquid AI's d1-3B is built from LFM2.5-VL-3B and…
X: The classic hotdog classifier as a real-time vision decision. Fayaz Ahmed rebuilt the not hotdog app on…
X: Liquid d1 with vision
More Jev resources like this
The experimental image-input reference: PNG or JPEG files in a top-level assets map, referenced by string anywhere in state, instructions, or criteria, behind the TypeSafe-Danger-Zone-Images header.
Jev Practices & PatternsDocs#official#docs#vision
Official
A maintained list of jev-1.13's known failure modes, literal reading, unreliable counting, dates as text, indirection, context rot, and contradictory criteria, each with a guardrail.
Jev Practices & PatternsDocs#official#docs#evaluation
Official
Explore TypeSafe use cases by industry and turn promising ideas into concrete software workflows.
Jev Practices & PatternsDocs#official#use-cases
Official
Documents a failure mode in letter-bound remote sensing classification: 97.71% OA on UCM in vocabulary but 7.35%–96.40% cross-dataset accuracy and inverted confidence (AUROC 0.38). RSAJEV replaces the letter readout with training-free prototype retrieval, reaching 89.43% on the UCM cross-dataset gate.
Jev Practices & PatternsArticle#community#research#arxiv
Community
Reference-bound Visual Jev rewards verify whether requested attributes, actions, and relations hold for specified reference subjects in multi-subject image generation. Training a Qwen3.5-4B verifier as a GRPO reward raises a GPT-5.4 composite score from 41.78 to 52.50 on a 897-task MICo-Bench subset.
Jev Practices & PatternsArticle#community#research#arxiv
Community
An evaluation of Jev and two open-weight typed decision models when option names and definitions conflict. With yes/no names, decision-flip rates are up to 70.4 percentage points higher than the 0/1 control, and mean AUC drops from 93.8% to a below-chance 23.2%.
Jev Practices & PatternsArticle#community#research#arxiv
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.