TypeLLM said Jev 1.14 cannot return every vehicle box, including Tesla's flying Roadster; this desk's every-vehicle Box was one rectangle, n = 3, and type array returned HTTP 400

Verify

Primary post is Diogo Almeida, @CompleteSkeptic, October 14, 2026, 21:52 UTC, status 2110488927329218563. The text names a new type in the gray area of decisions: a single bounding box, only on jev-preview. TypeSafe's recap, 01:02 UTC October 15, status 2110536690037391732, calls it the bounding box primitive for locating at the speed of Jev, and links docs.typesafe.ai/primitives/box. The TypeSafe tweet still is the mug diagram, left 0.44, right 0.84, top 0.25, bottom 0.82. The still on this page is TypeLLM's locate-vehicles cookbook at 14.00 s, six vehicles, the flying Roadster labeled 5.

Opened October 15, 2026, the Box page says a Box is a System One question type for finding where something is in an image. The answer is the four edges of a box around it, as fractions of the image's width and height. Dreamer-ready, same images header as the Images page. Box questions need Jev 1.14: send jev-preview or jev-1.14.0. type is always box. instructions names the target and the image. Box takes no criteria. left and right are fractions of width. top and bottom are fractions of height. Origin is the top-left. left is less than right, top is less than bottom. The answer includes confidence from 0 to 1. There is no not-found value. The docs tell the caller to ask a Noul whether the target is present and to ignore the box when that Noul is low. One box per question. On 36 test crops of paintings, four edge Nouls caught 30 of the cuts. A single is-all-of-it-inside question caught 26. Drawing the box as an outline caught 20. This desk did not rerun those 36 crops.

The blog calls Box the first type that is not purely meant for smart-if statements, enabled only for jev-1.14.0.

On October 15, 2026, Date header 01:22:15 GMT, a POST to https://api.typesafe.ai/v1/systemone, model jev-preview, header TypeSafe-Danger-Zone-Images: true, returned HTTP 200 and named jev-1.14.0. The image is a 400 by 400 PNG drawn here, red square from 0.40 to 0.80 of the width and 0.20 to 0.70 of the height. Box instructions: the red square in photo. Returned box left 0.4, right 0.8, top 0.16, bottom 0.7, confidence 0.93. The Noul on that same request, Is there a red square in photo, was 0.39. Usage 760 input tokens and 38 output tokens. Desk clock 457 ms. List-price arithmetic $0.000032.

Date header 01:22:16 GMT, the same Box question, model jev-1.13.0, images header on, returned HTTP 400. The message is Invalid request. Desk clock 424 ms. Key, request JSON, and request ids are not recorded here.

Date headers 06:16:36 GMT to 06:16:40 GMT, three more Box answers on that same PNG, model jev-1.14.0. Edges: 0.39/0.79/0.16/0.71 at 0.90, 0.40/0.80/0.16/0.71 at 0.92, 0.40/0.79/0.17/0.70 at 0.92. Mean left 0.40, right 0.79, top 0.16, bottom 0.71, confidence 0.91. The Nouls on those three requests are on the Images page. Date header 06:16:41 GMT, an empty 400 by 400 field still returned a box, top 0.02, bottom 0.98, left 0.01, right 0.98, confidence 0.05. The Noul on that request was 0.04. Date header 06:16:42 GMT, the same Box question on jev-1.13.0 returned HTTP 400, Invalid request.

TypeLLM posted at 05:47 UTC that Jev 1.14 introduces image bounding box generation and that TypeLLM now supports it through its API. The example page, opened October 15, prints a dog_head box top 0.352, bottom 0.652, left 0.344, right 0.628, time 0.48 s, cost $0.000053, 1,051 input tokens. This desk did not call TypeLLM.

brainstormity posted at 14:08 UTC that JEVsnap's optional visual check uses Jev Box. The README, opened later, says that path is off by default and that a downscaled copy is sent only with earlier hits already blacked out. This desk did not install the app.

Date headers 14:57:24 GMT to 14:57:25 GMT, three Box answers on a 400 by 400 PNG drawn here, blue square 0.08 to 0.28 of both axes, red square 0.40 to 0.80 of the width and 0.20 to 0.70 of the height. Instructions: the blue square in photo. Model jev-1.14.0. Edges: 0.07/0.30/0.14/0.29 at 0.94, 0.07/0.30/0.14/0.29 at 0.93, 0.07/0.30/0.14/0.30 at 0.94. Mean left 0.07, right 0.30, top 0.14, bottom 0.29, confidence 0.94. The Nouls on those three requests are on the Images page. Date header 14:57:28 GMT, the same Box question on jev-1.13.0 returned HTTP 400, Invalid request. Desk clock 601 ms. n = 3.

Aviz Maeir, @aviz85, posted at 22:04 UTC October 15, status 2110854390999306324, Jev found Waldo in 0.2 seconds. The thread is sixteen tweets. All numbers below are from his runs. Synthetic computer-use screens 13 of 13. Real-looking UIs 24 of 28, 86 percent, about 0.2 s and about $0.0001 each. Fifty-four symbol and drawing tests 50 of 54, 93 percent. Where's Waldo on scenes of 30 to 120 people 70 of 72, 97 percent. Crowds of 200 to 600 people, Waldo at one third to one half size, 52 percent in one call and 69 percent after 2x2 tiles. On those Waldo scenes, confidence 0.7 or above was right 79 of 79. Whack-a-Mole, three one-minute rounds: 31 of 33, 54 of 60, 51 of 59. About 170 ms per call. About 620 frames, about 4 cents. Forbidden bunny and bomb, hard round 25 of 30. All five mistakes were the bomb, each with box confidence below 0.3. Four box plus yes/no questions in one request billed 2,266 tokens against 8,635 as separate calls, about 3.8 times. Same 180 scenes against OpenAI Responses API plus strict JSON: Jev 1.14 82 percent, 0.21 s, $0.15 per 1k. gpt-6-luna 84 percent, 1.5 s, $0.22 per 1k. gpt-6.1-sol 95 percent, 3.4 s, $5.03 per 1k. Confidence 0.7 or above was right 107 of 108. Luna printed 0.9 to 1.0 on 27 confident misses. Dense crowds with look-alike decoys: sol 83 percent against Jev 52 percent. OpenAI's Decisions API cannot return boxes. This desk did not rerun the 180.

Kerry Lu and Mateo Wang's LiteLLM post, dated October 15, 2026, adds day-0 support for box on POST /v1/systemone and POST /v1/decisions. The proxy forwards client headers that start with TypeSafe-Danger-Zone- to TypeSafe only. The decisions docs print an illustration: noul 0.99 and box top 0.11, bottom 0.51, left 0.49, right 0.83, confidence 0.94, 756 input tokens. That JSON is a worked example. This desk did not run the proxy.

dave, @dave8172, posted at 23:14 UTC, Where's Pip. The play URL is quirkyagents.com/wizard/projects/wheres-pip/play/. This desk did not play it.

Date headers 00:28:46 GMT to 00:28:53 GMT October 16, 2026, this desk sent 16 POSTs to https://api.typesafe.ai/v1/systemone plus GET /v1/models. Fourteen POSTs printed usage, 24,552 input tokens and 674 output tokens. List-price arithmetic $0.001031. GET at 00:29:34 GMT named jev-latest and jev-preview, both release_date still 2026-09-10. Every jev-preview POST named jev-1.14.0.

Three POSTs, Date headers 00:28:47 GMT to 00:28:48 GMT, used a 400 by 400 PNG drawn here: a Checkout panel and a green Continue button from 0.55 to 0.90 of the width and 0.70 to 0.88 of the height. Instructions: the green Continue button in photo. Model jev-1.14.0. Edges 0.55/0.91/0.71/0.88 at 0.92, 0.55/0.91/0.71/0.88 at 0.90, 0.55/0.90/0.71/0.88 at 0.91. Mean left 0.55, right 0.91, top 0.71, bottom 0.88, confidence 0.91. IoU 0.92, 0.92, 0.94, mean 0.93. The drawn button's center sat inside all three boxes. Noul Is there a green Continue button in photo returned 0.98, 0.99, 0.99. Each used 1,655 input tokens and 36 output tokens. Desk clocks 413 ms, 416 ms, 480 ms. n = 3.

Date header 00:28:49 GMT, the same Box on jev-1.13.0 returned HTTP 400. Date header 00:30:08 GMT, a second POST printed Invalid request. A POST with the image and no TypeSafe-Danger-Zone-Images header, Date header 00:28:49 GMT, returned HTTP 400. Date header 00:30:08 GMT, that message is Images are experimental. Send the header TypeSafe-Danger-Zone-Images: true to use them.

Date header 00:28:50 GMT, one POST asked four Nouls and four Boxes on a 400 by 400 PNG with a red, blue, green, and yellow square, image in state once. Named jev-1.14.0. All four Nouls 1.0. IoU 0.76, 0.79, 0.83, 0.83. Centers inside all four. Confidence 0.88, 0.49, 0.75, 0.28. Usage 4,458 input tokens and 138 output tokens. Desk clock 432 ms. Four matching separate POSTs, one color each, Date headers 00:28:50 GMT to 00:28:51 GMT, used 1,653 input tokens each, sum 6,612. n = 1 for the combined call.

Date headers 00:28:52 GMT to 00:28:53 GMT, three POSTs on a 400 by 400 PNG with eighty gray circles and one red circle drawn at 0.60 to 0.74 and 0.36 to 0.50. Instructions: the red circle in photo. Named jev-1.14.0. Edges 0.59/0.72/0.36/0.48 at 0.78, 0.59/0.73/0.36/0.48 at 0.83, 0.59/0.71/0.36/0.48 at 0.84. Mean left 0.59, right 0.72, top 0.36, bottom 0.48, confidence 0.82. IoU 0.69, 0.75, 0.63, mean 0.69. Centers inside all three. Noul 0.99, 0.99, 0.98. Each used 1,653 input tokens. n = 3.

Date header 00:28:53 GMT, an empty 400 by 400 field still returned a box, top 0.02, bottom 0.98, left 0.02, right 0.98, confidence 0.04. Noul 0.02. Usage 1,653 input tokens. n = 1.

TypeLLM posted at 05:44 UTC October 16, status 2110970024328564741, that Jev 1.14 cannot detect every vehicle in an image and return an array of bounding boxes, including Tesla's new flying Roadster. The cookbook, opened the same hour, typellm.ai/cookbook/locate-vehicles-in-a-video, takes seconds 12 to 16 of a car show at eight frames a second. The question type is array, items type box. Instructions: Every vehicle, each on its own: every car, the truck, and any vehicle in the air. At 13.75 seconds the printed list is six boxes, the last at top 0.493, bottom 0.597, left 0.699, right 0.871. The table: 12.00 to 13.25 five vehicles on the ground; 13.38 to 15.25 six, with a sixth in the air; 15.38 to 16.12 five after it has flown out. Each frame 2,096 input tokens, about $0.0001. Four seconds under half a cent. A frame with no vehicle returns []. The notebook asks for typellm>=0.7.2. This desk did not call TypeLLM.

Elon Musk posted at 01:57 UTC October 16, status 2110913107094606330, Turns out the future does have flying cars. The clip is 15.45 s. Tesla posted the Roadster livestream at 17:41 UTC October 15, status 2110788059008819215.

TypeLLM's 07:56 UTC reply, status 2111003375303336430, says Jev only supports one box per question, links docs.typesafe.ai/primitives/box, and says there is no way it can count and generate all the boxes of the same category. The 08:00 UTC clarification, status 2111004357378613761, repeats that Jev supports one bounding box per question and that the questions are independent. nathan L. at 07:42 UTC quoted Nathan LeClaire's 20:10 UTC October 15 post that Box is Yolo for everything.

The Box docs, reopened October 16, still say one box per question. To find several different things, ask one Box question for each in the same request. When the image might hold more than one of the same thing, describe the one you want, or add a Choice with none, one, and several. TypeLLM's output-types page, opened the same hour, documents arrays of boxes, from none to maxItems, at most 50, and [] when there is no match.

Date headers 08:19:43 GMT to 08:19:46 GMT October 16, 2026, this desk sent 6 POSTs to https://api.typesafe.ai/v1/systemone plus GET /v1/models. Five POSTs printed usage, 5,187 input tokens and 364 output tokens. List-price arithmetic $0.000218. GET at 08:19:43 GMT named jev-latest and jev-preview, both release_date still 2026-09-10. Every jev-preview POST named jev-1.14.0.

Three POSTs, Date headers 08:19:44 GMT to 08:19:45 GMT, used a 400 by 400 PNG drawn here: a red car on the ground at 0.08 to 0.32 of the width and 0.67 to 0.865 of the height, a blue car at 0.62 to 0.88 and 0.67 to 0.865, and a white vehicle in the air at 0.38 to 0.62 and 0.12 to 0.32. Instructions: every vehicle in photo. Model jev-1.14.0. Each answer was one box. Edges left 0.37, right 0.63, top 0.12, bottom 0.90, confidence 0.19; then 0.36, 0.63, 0.12, 0.89, 0.17; then 0.37, 0.63, 0.13, 0.34, 0.13. n = 3. Desk clocks 411 ms, 506 ms, 456 ms. Input tokens 759, 762, 758.

Date header 08:19:45 GMT, one POST asked a count Choice, a none/one/several Choice, a Noul, and three named Boxes on that PNG. Count chose three at 1.0. Several chose several at 0.97, several 0.98. In-air Noul 0.90. Red left 0.07, right 0.32, top 0.66, bottom 0.86, confidence 0.62, IoU 0.89. Blue 0.62, 0.89, 0.67, 0.85 at 0.25, IoU 0.89. Flying 0.38, 0.62, 0.12, 0.30 at 0.93, IoU 0.90. Usage 984 input tokens and 152 output tokens. Desk clock 612 ms. n = 1.

Date header 08:19:46 GMT, a POST with type array, items type box, instructions Every vehicle, each on its own, returned HTTP 400. The message is Invalid request. error_type api_usage_error. Desk clock 408 ms. No usage field.

Date header 08:19:46 GMT, one POST used a 960 by 540 JPEG taken from Elon Musk's 15.45 s video at about 7.0 s, the white Roadster in the air over a row of five vehicles on the ground. Named jev-1.14.0. Count chose six at 0.59, six 0.69, seven_or_more 0.27, five 0.04. Several chose several at 1.0. In-air Noul 0.94. Every vehicle one box left 0.30, right 0.68, top 0.55, bottom 0.69, confidence 0.22. Flying 0.58, 0.70, 0.36, 0.49 at 0.69. White van 0.46, 0.54, 0.52, 0.69 at 0.84. Usage 1,924 input tokens and 152 output tokens. Desk clock 555 ms. n = 1. Key, request JSON, and request ids are not recorded here.

Compare

Choice, Score, and Noul stay the three production primitives on /jev/. Box is a fourth type on jev-1.14.0. OpenAI Decisions documents predicate, choice, and score. It has no Box on the pages this desk opened.

imajev-4b returns Choice, Score, or Noul from up to two photos. It does not print a bounding-box field on the imajev-4b page.

The mug diagram on TypeSafe's tweet is an illustration, left 0.44 and right 0.84. This desk's red square was drawn at 0.40 to 0.80 and 0.20 to 0.70. The returned top 0.16 sits 0.04 above the drawn top. n = 1. The 36-crop edge-question table is TypeSafe's, and this desk did not rerun it.

The 06:16 UTC three-call mean, left 0.40 right 0.79 top 0.16 bottom 0.71, is a later sample of the same PNG. The empty-field box at confidence 0.05 is n = 1. TypeLLM's dog_head example is a different host.

The 14:57 UTC three-call mean, left 0.07 right 0.30 top 0.14 bottom 0.29, is a later sample on a two-square PNG. Drawn blue edges are 0.08 / 0.28 / 0.08 / 0.28. Do not merge 0.94 with the 06:16 red-square 0.91. JEVsnap is a wrapper. Jev4Mellea does not send Box.

Aviz's 82 percent is 180 scenes against gpt-6-luna 84 percent and gpt-6.1-sol 95 percent on his Responses JSON harness. It is not this desk's Continue IoU 0.93, not TypeSafe's 30 of 36 crops, and not OpenAI Decisions, which has no Box. 107 of 108 is his confidence cut at 0.7. This desk's Continue and Waldo calls at 0.7 or above were right 8 of 8. Combined yellow at 0.28 and blue at 0.49 were also right. Those eight plus two are n = 10 on two synthetic PNGs.

4,458 against 6,612 is this desk's four-color PNG with the image in state. Aviz's 2,266 against 8,635 is a different file and a different question layout. Do not merge 3.8 times with 6,612 / 4,458. 70 of 72 is his 30 to 120 person Waldo set. This desk's red circle among eighty gray circles is n = 3 on one PNG.

TypeLLM's cookbook prints six boxes at 14.00 s, including the car in the air. That host returns an array. This desk's every-vehicle Box on the three-car PNG returned one rectangle, n = 3. A type array request returned HTTP 400. Named Boxes in one request returned three rectangles. Choice counted three at 1.0 on that PNG and six at 0.59 on Elon's frame. Do not merge TypeLLM's six boxes with this desk's one rectangle. Do not merge 0.59 with Aviz 82 percent or Continue IoU 0.93.