Independent. Not affiliated with TypeSafe AI, the maker of Jev. Disclosures

Decision LayerUnofficial

Try “routing”, “confidence”, “leads” or “noul”.

    Guides · field guide to Jev

    Jev decides. We show you what to let it decide.

    Jev is TypeSafe AI’s new kind of model: it reads a ticket, an email or a transcript and returns typed answers with probabilities, not text. We explain the use cases, write the how-tos, run the tests, and tell you where it breaks.

    Browse use cases Start here
    $0.042per million input tokens; output free (docs)
    3question types: Choice, Score, Noul (docs)
    Sep 152026 early-access launch (The Register)
    POST /v1/systemone · jev-latest

    state

    “Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP.”

    questions

    • departmentChoicebilling · technical · sales
    • frustrationScorecalm → frustrated → very angry
    • is_urgentNoulyes / no

    answers

    department
    billingtechnicalsales
    frustration1.0
    calmfrustrated but civilvery angry
    is_urgent1.0

    → route to technical, flag urgent

    Example request and answers from TypeSafe’s quick start, shown as published. Not our test.

    Choice Pick one of a set

    Which team, which product, which next step. Up to 255 options, with a probability for each.

    billing · technical · sales

    Score Place on a scale

    How urgent, how good a fit, how complete. You describe every level in words.

    cold → warm → hot

    Noul Yes or no, as a probability

    Did the rep read the disclosure? Is this a refund request? One number from 0 to 1.

    p(yes) = 0.94 (illustrative)

    The idea that makes it useful

    Let code act when Jev is sure. Send the rest to a person.

    Every answer carries a probability. You decide the lines. TypeSafe’s own routing example sends anything under 0.5 confidence to a human; the right lines for you come from testing on your own data.

    Read: setting confidence thresholds

    Illustration. Where the lines fall is your call, set from your own tests.

    01Use cases for insurance agencies and contact centers

    All agency use cases
    SupportStrong fit

    Support ticket routing

    Which team gets the ticket, how urgent it is, how upset the customer sounds.

    ChoiceNoulScore
    Guide in progress
    InsuranceTest first

    Insurance lead scoring

    How well an inbound lead fits, what they are shopping for, and whether they are ready to talk now.

    ScoreChoiceNoul
    Guide in progress
    InsurancePerson in the loop

    Claims intake triage

    Claim type, how complex it looks, and whether it should go to a senior adjuster for review.

    ChoiceScoreNoul
    Guide in progress
    SupportTest first

    Call QA scorecards

    Did the rep cover each required point: greeting, disclosures, needs questions, next step.

    NoulScore
    Guide in progress

    02Beyond insurance

    All 12 use cases
    AI engineeringTest first

    Guardrails for AI agents

    Is this command, tool call or output safe to run, or should it stop for review?

    NoulChoice
    Guide in progress
    AI engineeringTest first

    Model routing

    Is this request easy enough for a cheap model, or does it need a stronger one?

    ChoiceScore
    Guide in progress
    AI engineeringStrong fit

    Re-ranking search results

    Which of these passages actually answers the question.

    ScoreChoice
    Guide in progress
    OperationsPerson in the loop

    Candidate screening

    Which applicants make the shortlist for a person to review.

    Score
    Guide in progress

    03How-tos

    All 10
    1. 01Make your first Jev callAn API key, one request with curl, then the same request in Python and JavaScript.Start
    2. 02Call Jev through Vercel, OpenRouter or CloudflareThe same question on each gateway, and what changes (Vercel calls the yes/no type “boolean”).Start
    3. 03Pick the right question typeChoice for a set of options, Score for a scale, Noul for yes or no, with a test for each.Start
    4. 04Write questions Jev answers wellNarrow, one-property questions instead of one broad one, and why that matters.Build
    5. 05Shape the state you sendOnly what the decision needs, as structured JSON, with numbers and dates turned into words.Build

    04From the blog

    Blog
    1. 01Text written to fool it: building a claims fast track that a claimant can’t talk their way intoJev reads everything you send it as evidence, including a line a claimant wrote to steer it. Here is what new research says that line can and can’t do, and how to design a fast track for small glass claims so it doesn’t matter.Worked example
    2. 02Jev, its clones and your service inbox: what’s actually under the hoodDecision models are multiplying. Here is what Jev is likely built on, how it differs from a chatbot and from a classifier you train yourself, and why the test you run matters more than the brand. Shown on a made-up email to an insurance agency.Comparison
    3. 03A function call for underwriters: sorting commercial submissions with JevWhat “unstructured state in, typed probabilistic decisions out” means in practice, shown on a made-up contractor submission at an insurance agency.Worked example
    4. 04Labels, not essays: sorting first notice of loss reports with JevA worked example for a claims desk. Four narrow questions, a few lines of routing code, and a person on every decision that touches a payout.Worked example
    5. 05What Jev is, and what it isn’tA plain-English guide to TypeSafe AI’s decision model: what it hands back, why developers like it, and the five things it is bad at.Explainer

    Reality check

    We read the claims closely so you don’t have to.

    Jev is new, fast and cheap, and the marketing is loud. We label every vendor number as a vendor number, test what we can, and say plainly when a claim needs a footnote.

    How we test and source
    “Zero hallucinations”
    Means answers always fit the format you asked for. TypeSafe’s launch post calls the 0% figure “not empirical”. Jev can still pick the wrong answer.
    “193.6x faster”
    The homepage example beside it (0.114s vs 8.566s) works out to about 75x. Both come from TypeSafe’s own evals.
    Confidence
    The docs define it as the shape of the probabilities and make no claim it equals real accuracy. Test it on your data.

    05Start where you are

    I run a team

    Support, sales or operations. You want to know what Jev can decide for you and what it costs.

    I write code

    You want working requests, patterns that hold up, and honest test numbers.

    I run an insurance agency

    Leads, calls, service tickets and claims intake, with a person in the loop.