Start hereExplainer, checked 28 September 2026 · model facts rechecked 16 October 2026, 05:21 UTC · model IDs and aliases rechecked 24 October 2026, 05:12 UTC · question types (Box on 1.14) read 7 November 2026, 15:29 UTC

What is Jev (TypeSafe AI)? A plain explanation of Choice, Score and Noul

Jev is a hosted model from TypeSafe AI. Your application sends it some text or structured data (the state) and one or more typed questions. For each question Jev returns an answer from a fixed set, with probabilities. It does not write prose or code.

Short answer

Jev is a decision model, not a chatbot. You ask "which of these options?", "where on this 2–10 level scale?" or "yes or no?", and it returns the answer and its probabilities. Your code then checks the answer and decides whether to act, send the case to a person, or fall back to another method. Documented TypeSafe API and primitives docs, checked 28 Sep 2026

Name check: this page is about TypeSafe AI's Jev model. It is not about Japanese encephalitis virus (JEV) or any person named Jev.

One request, step by step

1. Your application sends
model: jev-1.13.0
state: the ticket text
questions: typed Choice, Score or Noul
POST /v1/systemone with a Bearer key
2. Jev returns, per question
An answer from the fixed set you gave it
Probabilities for that answer set
No free text and no action
3. Your code decidesAct: a clear, allowed answerHold: send to a personFall back: a rule or another model

One request in three steps. Field names come from the TypeSafe API reference (checked 28 Sep 2026); the exact JSON body is on the Python example page. A valid answer can still be wrong, so step 3 is always your code.

The request and response, annotated

Documented Request adapted from TypeSafe's quickstart (model pinned, one question kept). Not executed by the pool.

curl -X POST https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Help! My payouts have been failing for 3 days.",
    "model": "jev-1.13.0",
    "questions": {
      "department": {"type": "choice", "instructions": "Which team should handle this?",
                     "criteria": {"billing": "Payments, invoicing, refunds",
                                  "technical": "Bugs, outages, integrations",
                                  "sales": "Pricing, upgrades, new accounts"}}
    }
  }'

Documented Response fields and values from the API reference's example for a department Choice. The two examples come from different documentation pages, so this response is not the output for the request above.

{
  "model": "jev-1.13.0",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "probabilities": {"billing": 0.88, "technical": 0.12, "sales": 0.0},
      "confidence": 0.81
    }
  },
  "usage": {"input_tokens": 318, "output_tokens": 34}
}

model is the versioned model that answered, even if you sent an alias such as jev-latest.
choice is the highest-probability option; probabilities sum to 1.
confidence (0.81) measures how peaked the probabilities are. It is not the chance that "billing" is right.
usage.input_tokens (318) is what TypeSafe bills; output tokens are reported but free on the direct route. See cost per decision.

Checked 28 Sep 2026. Full, version-pinned Python code is on Call Jev from Python.

The question types: Choice, Score, Noul (and Box on 1.14)

Type	You provide	Jev returns	Use it for	Watch out for
Choice	Up to 255 named options	The chosen option, a probability for each option, and a `confidence` number	Routing to a queue, picking a label, picking a UI element or tool	Choice always picks something. Add an explicit "other" or "review" option when the list may be incomplete.
Score	An ordered rubric: at least 2 levels recommended, at most 10 accepted	A probability-weighted position on that scale, probabilities per level, and a `confidence` number	Urgency, severity, toxicity, quality grades	The number is a position on your rubric, not a calibrated percentage.
Noul	A yes/no question	The probability of yes	Gates: "is this spam?", "does this quote support the claim?"	There is no third "unsure" outcome and no separate confidence field; you choose the cut-offs.

Documented Option and level limits from TypeSafe primitives and the API reference, checked 28 Sep 2026. confidence measures how peaked the probabilities are; it is not the probability that the answer is correct. Details are on Confidence thresholds.

Choice, Score and Noul work on both versions. Box (a bounding box in an image: "Where is it in this image?") is a fourth question type on Jev 1.14 only, behind the images opt-in header; TypeSafe's images page says images and Box questions "aren't yet up to our standards to call ready for production". Documented Models page "Question types" row ("Box questions need Jev 1.14") and quickstart, read 7 Nov 2026 (R27-S05, R27-S03); images page (R27-S06). Box was not tested here. Change, 7 Nov 2026: this section was titled "The three question types"; the Box line is new (first observed on the models page 7 Nov 2026).

Worked example: a support ticket

A customer writes "I was charged twice this month and can't log in to see the invoice." Your application could ask two questions in one request:

Question	Type and answer set	What your code does with the answer
Which queue?	Choice: billing, bug, account, review	Routes to the chosen queue only if its probability is above the cut-off you tested; otherwise sends it to "review".
How urgent?	Score: 1 (low) to 4 (outage)	Sets a priority field. Refunds and account changes are never triggered by the answer alone.

This ticket mentions two problems (billing and login). A good design lets it land in "review" or in both queues, rather than forcing one label. This example restates the support-routing pattern in TypeSafe's and Netlify's documentation; it is not a result of a call made by this pool. What five real ticket-triage projects do when Jev is unsure or fails is on support-ticket triage.

Current model facts

Model IDs: jev-1.14.0 (Jev 1.14, available from 14 Oct 2026; released as a preview, now jev-latest) and jev-1.13.0 (Jev 1.13, by ID) Documented 24 Oct 2026, 05:12 UTC; both first listed together on 16 Oct 2026, 05:21 UTC
Aliases: jev-latest now gives jev-1.14.0, and jev-preview also gives jev-1.14.0; TypeSafe says "There is no preview build available right now" (TypeSafe models page, checked 24 Oct 2026, 05:12 UTC). TypeSafe published no dated announcement of the move; jev-1.13.0 is still listed and callable by its ID, with no deprecation date. Before the move: on 16 Oct 2026, 05:21 UTC, jev-latest pointed to jev-1.13.0 and jev-preview to jev-1.14.0. TypeSafe said on 14 Oct 2026 that jev-latest would move to jev-1.14.0 "in a week"; on 23 Oct 2026, 05:45 UTC, it still pointed to jev-1.13.0, and TypeSafe had published no date for the move (TypeSafe blog post, 14 Oct 2026; models page, 23 Oct 2026). Aliases can move; pin a versioned ID if you tuned thresholds. Earlier: both aliases pointed to 1.13.0 on 27 Sep 2026.
Listed price: $0.042 per million input tokens; output free; the same for both versions (16 Oct 2026, 05:21 UTC; unchanged since 28 Sep 2026)
Access: Signups open since 27 Sep 2026; no free credit for new users (status)
Request limits: 64k tokens per request and 32k for state plus the longest question by default, on both versions; 262k with the opt-in TypeSafe-Danger-Zone-Long-Context header, which is experimental (16 Oct 2026, 05:21 UTC)
Input: Text or structured data by default. Images only with the opt-in TypeSafe-Danger-Zone-Images header: experimental, and not production-ready per TypeSafe. No audio or video (16 Oct 2026, 05:21–05:22 UTC)
Weights: TypeSafe's Jev weights were not found published (checked 26 Sep 2026)

Sources: TypeSafe models page and API reference, rechecked 28 Sep 2026 (release 1 report). Route prices are compared on Where to use Jev. Model IDs, aliases, price, request limits and input rechecked on the models page and images page, 16 Oct 2026, 05:21–05:22 UTC (R17-S05, R17-S06); the alias move was TypeSafe's statement in its blog post of 14 Oct 2026 (R17-S09) and was first seen on the models page on 24 Oct 2026, 05:12 UTC (R23-S05). TypeSafe describes 1.14 as "small, subtle improvements to everything", with "the largest gain in multi-lingual support" Reported (vendor claim, not tested here).

Change, 24 Oct 2026: the model ID line said 1.14 was "released as a preview" and the alias line said jev-latest points to jev-1.13.0 (16 Oct check, re-dated 23 Oct). They now give the 24 Oct check: both aliases on jev-1.14.0, jev-1.13.0 by its ID with no deprecation date (R23-S05, R23-S08, R23-S09).

Change, 23 Oct 2026: the alias line now gives the 23 Oct check (jev-latest still on jev-1.13.0; no date published for the move); it said TypeSafe "says" the alias will move "in a week" (R22-S05, R22-S09).

Change, 16 Oct 2026: this list showed one model ID (jev-1.13.0), both aliases on 1.13.0, 64k/32k as the only request limits, and "no raw images". TypeSafe added Jev 1.14 on 14 Oct 2026 as a preview, with experimental opt-ins for image input and 262k long context on both versions.

What Jev does not do

Jev returns probabilities over options you define; it does not generate text, take audio or video (images: since at least 7 Nov 2026 the models page lists "Text, images" as the input of both versions, while the Images page still calls them an opt-in not yet ready for production; an experimental opt-in since 14 Oct 2026), do arithmetic or date comparison reliably, or guarantee that a valid answer is correct. TypeSafe's model jaggedness page (last reviewed 17 Sep 2026) lists the weak task types with workarounds; all 17 documented limits are on Limitations and fit check. Documented checked 28 Sep 2026

What was not verified

This pool has not made a Jev call; the example above is a design, not an output. No image or long-context request was sent.
Accuracy on any task is not established here. See Confidence thresholds for how to test on your own data.
Independent open-weight Jev-like models (Kev, CLM, Laya and others) are not TypeSafe's weights and have no verified parity; see Alternatives.