Do questions see each other's answers?

Official docs say questions in one request are independent. If a later judgment needs an earlier answer, make a second request in code.

TypeSafe says questions run in parallel against a state that is ingested once. Adding questions barely changes latency and costs only the extra question tokens.

Is calibration a guarantee on one answer?

No. Official docs say calibration is measured across groups of predictions and does not guarantee that an individual answer is correct.

How Jev Works | Jev AI Hub

How TypeSafe Jev works: architecture, parallel sampler, and how one state is evaluated against typed questions.

Jev's runtime story is short: state in, typed answers out, code stays in control.

The loop

Your application gathers only the fields the decision needs.
You write one snap judgment per question.
You POST to https://api.typesafe.ai/v1/systemone.
You branch, threshold, or escalate using the returned values.

state + questions
        │
        ▼
   one HTTP request
        │
        ▼
  Jev evaluates each
  question in parallel
        │
        ▼
 typed answers + probabilities
        │
        ▼
     your code

State

State is the evidence. Official docs accept:

a plain string
a JSON object
an array of text

Images, audio, and video are not supported. Point a question at a field with a backticked path such as `ticket.messages[0].text`.

Send less than you think. Official jaggedness notes say unrelated detail acts as a distractor.

Questions are atomic

A good question is something a knowledgeable person could decide in a few seconds given the right excerpt.

Bad: “Analyze this ticket and decide the best course of action.”

Better: three questions — department (Choice), urgency (Noul), frustration (Score) — then a formula in code.

Official primitives docs recommend splitting multi-factor judgments and combining answers yourself. When priorities change, you edit a weight, not a prompt novel.

Jev architecture

TypeSafe's launch post describes a stack built for automation, not chat: a new model architecture, a parallel sampler, and RLCD (Reinforcement Learning for Calibrated Decisions).

What you can use without a paper:

One state is ingested once.
Every question in the request is evaluated against that state.
Answers come back as typed values plus probabilities, not a generated paragraph.

TypeSafe has not published weights or a methods paper on the pages we verified. Treat “Jev architecture” as that public stack description, not a reverse-engineered diagram. For “Jev benchmark” claims (speedups, workflow evals), read TypeSafe's launch post and their eval site; we do not reprint vendor charts as our measurements.

Parallel evaluation

Jev ingests the state once. Every question in the map is evaluated against that state. Official docs say adding questions barely changes response time and does not create “context-rot” between questions.

That leads to a practical pattern TypeSafe calls speculative fan-out: ask questions you might need, ignore the ones that do not apply. A cookbook on their site reports batching 13 questions in one call as cheaper and faster than 13 separate calls. Treat those multiples as a vendor-reported cookbook result, not a number we measured.

Answers your code can trust for shape

A successful Choice cannot invent a fourth department. A Score stays on your rubric. A Noul is a number in [0, 1].

That is a schema guarantee, not a correctness guarantee. Wrong labels still happen. Confidence and your own tests are how you decide when to act.

Probability versus confidence

Signal	Where	Meaning
`probabilities`	Choice, Score	Mass on each option or level
`confidence`	Choice, Score	How peaked that distribution is
`noul`	Noul	Probability the statement is true; no separate confidence field

Official confidence docs describe a three-path pattern: high → act, medium → confirm, low → escalate. Example floors in those docs (0.5, 0.9) are examples, not defaults. Tune them on your data.

Context budgets

When last verified, official models docs listed:

64k tokens for state plus all questions
32k tokens for state plus the single longest question

Stay under both. If a question is huge, shrink the question, not only the state.

When to use this workflow

Use it for production routing, filtering, and scoring where you want milliseconds and a closed answer space.

When not to use this workflow

If the next token must be prose, call an LLM. If the next token must be a sum, write a function. If one question depends on fetching new records, make two Jev calls.

Common mistakes

One mega-question that hides four decisions.
Tuning thresholds on jev-latest, then being surprised when the alias moves. Log response.model and pin jev-1.13.0 (or whatever version you tested) after calibration.
Assuming two phrasings of the “same” question are interchangeable. Official jaggedness notes show Noul and yes/no Choice on similar wording are not arithmetic inverses.

Continue with the API guide or Getting Started.

How Jev Works

Quick answer