Jev โ
Endpoint: https://mmu.vod-qcloud.com/v1/systemoneAuthentication: VOD ApiToken / Bearer
Jev is a System One decision model. It does not generate prose or chat text. Instead, it takes a piece of state (state) plus a set of structured questions and returns typed answers with calibrated probabilities that your code can branch on directly. It is built to replace the "ask a generative model for JSON โ parse โ extract a label" step in a pipeline โ it is not a substitute for a chat model.
The name System One refers to fast, intuitive judgement, as opposed to generative models that reason in multiple steps: Jev acts as a fast decision function inside your workflow โ unstructured state (text or images) in, typed decisions ready for code branching out.
Basic Information โ
| Item | Value |
|---|---|
| Endpoint | https://mmu.vod-qcloud.com/v1/systemone |
| Authentication | ApiToken issued by the VOD application / Bearer |
| Request | POST; JSON for text, multipart/form-data recommended for images |
| Output | Typed JSON (answer + calibrated probability); no generated text |
| Billing | Input tokens only โ output tokens are free |
Callable Models โ
| Model ID | Notes |
|---|---|
jev-1.14.0 | Adds image input and Box questions (experimental) |
jev-1.13.0 | Text only; returns 400 for Box questions |
jev-latest | Points to the latest version. Pin a specific version in production |
Question Types โ
Up to 8 questions per request. Questions are evaluated independently and share the input cost of the same state.
| Type | Purpose | Returns |
|---|---|---|
choice | Pick one option from a predefined set (classification / routing) | Selected option, probability distribution over options, confidence |
noul | Decide whether a statement holds (yes / no) | A probability between 0 and 1 (the probability of "yes") |
score | Rate on an ordered scale | Probability-weighted score, scale legend, distribution, confidence |
box (1.14+, experimental) | Locate a target in an image | Bounding box top / bottom / left / right (0โ1 ratios of image height/width), confidence |
- Choice: 2โ20 options, each with a short description.
- Score: 2โ10 levels. The returned score is a probability-weighted expectation and may fall between two levels.
- Noul: no separate confidence field โ threshold on the probability itself.
- Box: only answers "where", never "whether". It returns four bounds even when the target is not in the image, so it must be paired with a Noul question in the same request to decide whether the target exists.
Input Limits โ
Text โ
| Item | Limit |
|---|---|
state (serialized) | 8,000 characters |
instructions per question | 1,800 characters |
criteria (serialized) | 2,000 characters |
| Standard request total | 64K tokens; state + the single longest question โค 32K tokens |
| Long context (experimental) | Up to 262K tokens via a dedicated request header |
Images (1.14+, experimental) โ
| Item | Spec |
|---|---|
| Allowed in | state, instructions or criteria |
| Formats | PNG, JPEG (GIF, WebP and animated PNG are rejected) |
| Dimensions | 256โ1568 px per side; at least 78,400 total pixels |
| Size | โค 16 MiB per image; โค 32 MiB of images per request |
| Required header | TypeSafe-Danger-Zone-Images: true (when using images or Box) |
multipart/form-data is recommended for images: one JSON part named request, plus each image as its own file part. Base64 also works, but inflates image size by about one third and lowers throughput by about 14โ16% when upstream bandwidth is tight.
Jev rejects non-compliant images instead of resizing them โ validate dimensions and format on your server before uploading.
Output Characteristics โ
- Typed output: structured JSON your code can branch, sort and route on immediately โ no text parsing.
- Calibrated probabilities: trained for calibration, so confidence values are statistically meaningful and can drive auto-execute thresholds. Calibration is a population-level property; it does not guarantee any single answer is correct.
- No reasoning or explanation: only answers and probabilities are returned.
Typical Use Cases โ
- Classifying and routing support tickets
- Scoring urgency / risk level
- Safety checks on tool calls in agent workflows
- Violation-type decisions in content moderation
- Re-ranking search results
- Detecting and locating targets in images (with Noul for existence)
Request Examples โ
Ticket classification and urgency scoring (text, multiple questions) โ
{
"model": "jev-1.13.0",
"state": {
"message": "I was charged twice for my March invoice. Please refund the duplicate charge today.",
"customer_tier": "enterprise"
},
"questions": {
"category": {
"type": "choice",
"instructions": "What is this ticket about?",
"criteria": {
"billing": "Payments, invoices, or refund requests",
"technical": "Broken product or unexpected behavior",
"account": "Login, permissions, or profile issues",
"other": "Anything else"
}
},
"urgency": {
"type": "score",
"instructions": "How urgent is this ticket?",
"criteria": [
"Can wait for the next release",
"Should be resolved this week",
"Needs attention within 24 hours",
"Needs immediate attention"
]
},
"refund_requested": {
"type": "noul",
"instructions": "Is the customer explicitly requesting a refund?"
}
}
}Response (illustrative) โ
{
"category": { "choice": "billing", "confidence": 0.91 },
"urgency": { "score": 2.6, "confidence": 0.85 },
"refund_requested": { "noul": 0.96 }
}Safety gate before an agent tool call (Noul) โ
{
"model": "jev-1.13.0",
"state": { "action": "bash", "command": "rm -rf ./dist && npm run build" },
"questions": {
"allow_without_confirmation": {
"type": "noul",
"instructions": "Is this command safe to run without asking the user?"
}
}
}Suggested thresholds: โฅ 0.9 run directly; 0.6โ0.9 ask the user to confirm; < 0.6 block and escalate to a human.
Locating a target in an image (Box + Noul, multipart) โ
Scenario: decide whether a screenshot contains a "Submit" button and where it is. Because Box returns bounds even when the target is absent, a Noul question is required.
POST /v1/systemone
Content-Type: multipart/form-data; boundary=----JevBoundary
TypeSafe-Danger-Zone-Images: true
------JevBoundary
Content-Disposition: form-data; name="request"
Content-Type: application/json
{
"model": "jev-1.14.0",
"state": "Analyze the screenshot.",
"questions": {
"target_exists": {
"type": "noul",
"instructions": "Does the screenshot contain a 'Submit' button?"
},
"target_location": {
"type": "box",
"instructions": "the 'Submit' button in the screenshot"
}
}
}
------JevBoundary
Content-Disposition: form-data; name="screenshot"; filename="screen.png"
Content-Type: image/png
<raw PNG bytes>
------JevBoundary--- Images are sent as separate file parts. The JSON part needs no
assetsfield โ just mention the image by its reference name instate/instructions(the reference name is the file part'sname). - Write Box
instructionsas a description of the target, not a command โ e.g.the Submit button, notLocate the Submit button.
Response (illustrative) โ
{
"target_exists": { "noul": 0.97 },
"target_location": {
"box": { "top": 0.72, "bottom": 0.81, "left": 0.45, "right": 0.58 },
"confidence": 0.89
}
}In your code, check existence with the target_exists probability first, then convert the Box ratios to pixels:
if result["target_exists"]["noul"] > 0.8:
box = result["target_location"]["box"]
x = box["left"] * image_width
y = box["top"] * image_heightQuick check with curl โ
Text request (Noul)
curl -X POST https://mmu.vod-qcloud.com/v1/systemone \
-H "Authorization: Bearer $YOUR_API_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"model": "jev-latest",
"state": "Customer message: My checkout page shows a blank screen after I click Pay.",
"questions": {
"is_defect": {
"type": "noul",
"instructions": "Is the customer reporting a software defect?"
}
}
}'Image request (multipart, Box + Noul)
curl -X POST https://mmu.vod-qcloud.com/v1/systemone \
-H "Authorization: Bearer $YOUR_API_TOKEN" \
-H "TypeSafe-Danger-Zone-Images: true" \
-F 'request={
"model": "jev-1.14.0",
"state": "Analyze the uploaded screenshot.",
"questions": {
"target_exists": {
"type": "noul",
"instructions": "Does the screenshot contain a Submit button?"
},
"target_location": {
"type": "box",
"instructions": "the Submit button"
}
}
};type=application/json' \
-F 'screenshot=@screen.png;type=image/png'requestis a reserved part name and cannot be used as an image reference name.- The image part's
name(screenshotabove) must match the reference name used instate/instructions. - The image part's
Content-Typemust beimage/png,image/jpeg,image/jpgorapplication/octet-stream.
Integration Options โ
| Method | Notes |
|---|---|
| HTTP API | POST /v1/systemone; JSON for text, multipart recommended for images |
| Python / JS SDK | Handles multipart upload and image references automatically; recommended for production |
| Request header | Images and Box require TypeSafe-Danger-Zone-Images: true |
Known Limitations โ
Capability boundaries โ
Jev performs poorly on the following; use a generative model or a dedicated method instead:
- Cannot count: not suited to exact counting or numeric calculation.
- Cannot compare dates: ordering dates or computing intervals is unreliable.
- Lower accuracy outside English: English is the primary training language; test other languages on real data.
- Does not generate text: use a generative model for summaries, translation, rewriting or explanations.
- Weaker on multimodal than text: per the official guidance, all models including Jev do worse on multimodal data than on text โ break questions down to their simplest elements.
- Sensitive to adversarial content: user-controlled text in
statecan sway the decision; add safeguards around it. - Takes instructions literally: Jev answers the question you wrote, not the one you meant.
Structural limitations โ
| Limitation | Notes |
|---|---|
| No reasoning or explanation | Only answers and probabilities |
| Box has no "not found" answer | Returns bounds even when the target is absent; pair with a Noul in the same request |
| Box requires 1.14+ | 1.13 and earlier return 400 for Box questions |
| Images and Box are experimental | Request format, answers and limits may change between versions |
| Confidence โ correctness | High confidence does not guarantee a single answer is right |
| Unstable during testing | Availability and behavior may change without notice |
Not suitable for โ
- Open-ended chat, creative writing, long explanations
- Tasks needing deep reasoning or multi-step logic chains
- Highly regulated scenarios that require traceable reasoning
- A drop-in replacement for general chat or multimodal models
Recommendations โ
- PoC images and Box first: both are experimental โ validate with real data before going live.
- Always pair Box with Noul: never use Box alone to decide whether a target exists.
- Pre-process images: validate dimensions and format on the server before upload.
- Set confidence thresholds: higher thresholds for high-risk actions; route low confidence to a human.
- Add a fallback option to Choice: include
otherornone. - Keep questions atomic: split complex judgements into several questions and combine them in code.
- Pin the version: use a specific version in production (e.g.
jev-1.14.0), notjev-latest.