0:00Jev — TypeSafe AI's "System One" decision model — used to make decisions only on text. With Jev 1.14 (released Oct 14, 2026 as a preview), it can now look at an image and make fast decisions about what it shows, and it gains a new question type called box that returns the coordinates of an object. The pitch: instead of training a specialized vision model, you describe what to check, hand it an image, and get a decision. It's a beta feature available to all Jev users.
The 200-ticket demo
0:53The presenter's test: a small app that reads support tickets and checks whether each one includes a relevant screenshot to back up its claim (routing "no screenshot" tickets back to the user). On a public dataset of 200 tickets (from SWE-bench's bug reports), it classified all 200 images in about 4 seconds for ~2 US cents — with bounding boxes showing pass (image supports the claim), fail (problem not visibly clear), and a human review bucket where independent answers disagreed.
The box type (and its gotcha)
1:53The classic question types remain (e.g. null, the boolean-plus-probability); the new one is box. In a playground demo — a shelf of liquids — a "which item is drinkable?" null question returned 94% true, while a "which item matches this hazard pictogram?" box question pinpointed the grain alcohol (95% proof = flammable, unlike bleach). 2:57Asked "where does the last item sit on this pH scale?", it identified a lemon and selected pH 3 — using box coordinates rather than the other answer types. 3:13A claim-assessment demo (a shattered phone photo) scored "severe, not extreme" damage at 91% confidence / 93% probability. 4:01The gotcha to know: the box type always outputs a box — there's no "not found." If the target is completely absent, it still returns four edges, so the recommended pattern is to pair a box question with a null question that checks presence first, and ignore the box if the null says "not present."
Patterns
4:31Because you can drop images anywhere in the request (state, instructions, or criteria), a few patterns open up: compare against a known picture (is this the same person/product/place as the reference image?), define choices with images (criteria that are just memes, no text — "if the database is on fire, which meme?"), detect what changed (send the previous frame of security footage and ask "was there motion? an animal? a person arriving?"), and tiling for high-resolution images — cut into tiles, ask a choice question to pick the right tile, then send a higher-res crop of it. The reminder: it's not a generative text model; you have to give it options to choose.
The 1,000-ticket test
6:25To get calibrated numbers, they built 1,000 controlled cases with synthetic screenshots. It chewed through them at ~50 tickets/second (batches of 16), ~20 seconds total, 11 US cents. With a 65% confidence cutoff: 724 went back to the customer (wrong screenshot), 175 accepted, 101 sent for human review. 7:32Because it's synthetic, they know the right answers — so they can measure the three things that matter when deploying a decision model: accuracy (how often it's right), confidence (how strongly it favors its answer), and calibration (how often its probabilities match reality).
Accuracy vs calibration
8:36Accuracy: 93.4% (934/1,000 correct). The review gate: of the 899 decisions above the 65% threshold, Jev made zero mistakes — all 66 errors landed in the 101 sent for human review. Then calibration, the subtler one: the 899 auto-decisions carried an average probability of 95.3%, which predicts ~42 mistakes — but there were zero, so Jev was under-confident there. The 101 review cases averaged 48% probability but only 34% actual accuracy — over-confident. The takeaway: confidence is a usable gate (it cleanly routed every mistake into review), but the raw probabilities aren't yet calibrated enough to read as true error rates.
Limits & competitors
9:47Limitations: still images only (PNG/JPEG), 16 MB max image, 32 MB max request body, and resolution bounds — most of which you work around by resizing/cropping up front. Enabling the feature means setting a TypeSafe "danger-zone" images header. 10:14Jev isn't first to multimodal decision models: the video points to Cloudflare's Clef (27B, open-source, Jev-API-compatible) and to two open-source models covered in earlier videos — Win12B and Decider4B, both built on Qwen and Gemma bases.
Claims checked
Jev 1.14 is real. TypeSafe AI released it Oct 14, 2026 as jev-preview, adding experimental image input, the box question type, and a 262k-token long context (per the TypeSafe blog and MarkTechPost). The video's "released yesterday" and the box-type description match.
Cloudflare's competitor is Clef (not "Cleft" as auto-captioned) — a 27B multimodal decision model, open-source and Jev-API-compatible, on Workers AI.
All performance numbers are the presenter's own runs: 200 tickets in 4s for 2¢, 1,000 synthetic tickets at ~50/s for 11¢, 93.4% accuracy, zero mistakes above the 65% gate, and the calibration findings. These are his benchmark results, not independent measurements. Win12B and Decider4B are relayed as named.
The AI Automators community and its course are promoted at the end; no third-party sponsor. The "always outputs a box" gotcha and the null+box pairing are the presenter's recommendations, not vendor documentation.