JevKit

Jev explained

Official

What is Jev?

Jev is TypeSafe AI's flagship model and the first model in the category the company calls System One models. It turns application state and typed questions into structured, probabilistic decisions rather than primarily generating open-ended text.

Build a Decision SchemaRead the official docs

The JevKit builder is a generic teaching and planning tool. It does not generate an official Jev API payload.

Jev in one sentence

What exactly is Jev?

Jev is a TypeSafe AI model for fast, focused judgments: it evaluates developer-defined Choice, Score, and Noul questions against application state and returns typed answers with probabilities that code can inspect and act on.

From state to action

How Jev works

A Jev-shaped workflow separates model judgment from application policy. The model answers bounded questions; your code decides what those answers mean operationally.

  1. 01

    Application state

    The text or structured context your workflow already has.

  2. 02

    Typed questions

    Focused Choice, Score, and Noul judgments.

  3. 03

    Jev

    The System One model evaluates each question against the state.

  4. 04

    Typed decisions

    Values, probability distributions, and confidence where applicable.

  5. 05

    Code / action

    Your policy routes, verifies, asks for review, or takes an action.

Different task shapes

Jev vs a normal LLM

Jev is not a universal replacement for an LLM. They solve different shapes of problems: Jev is designed for bounded decisions inside software, while LLMs are built for flexible generation and broader reasoning.

DimensionJevGeneral-purpose LLM
Primary jobMake focused judgments for softwareGenerate text for people or tools
OutputTyped answers and probabilitiesGenerated tokens, usually text
Answer spaceDefined by the developer in advanceOpen-ended unless constrained by another layer
Typical integrationBranch, rank, route, verify, or gate in codeChat, write, explain, reason, or generate
UncertaintyProbabilities; Choice and Score also include confidenceVaries by model and integration
Best fitRepeated, bounded decisionsOpen-ended generation and broad reasoning

On narrow screens, scroll the comparison table locally.

TypeSafe terminology

What is a System One model?

System One model is TypeSafe AI's name for a class of models built to make fast, focused, structured judgments that software can use directly. Jev is the company's first model in that category.

The name emphasizes a different task shape from open-ended generation: evaluate a state, answer narrow questions, and return typed values and probabilities. “System One model” is TypeSafe AI's name for this model category, and JevKit uses the term in that TypeSafe-specific sense.

Three primitives

Choice, Score, and Noul

Each primitive asks one focused question and returns a shape your code can consume. The examples below are illustrative, not outputs from a live Jev call.

Choice

Official

Which one?

Selects one result from options defined by the developer.

Illustrative

Which team should handle this ticket?

Billing · Technical · Sales · Other

Score

Official

Where on this scale?

Places the state along an ordered scale whose levels you define.

Illustrative

How severe is this issue?

Low · Medium · High · Critical

Noul

Official

Is this true?

Makes a probabilistic yes/no judgment and returns the probability of yes.

Illustrative

Does this message request a refund?

0.94 probability of yes

Noul is not a plain Boolean. It is a probabilistic yes/no judgment: a value near 1 is a strong yes, near 0 is a strong no, and near 0.5 indicates uncertainty.

Support routing

Illustrative

A structured example

The useful unit is not a generated reply. It is a small set of judgments that code can combine with policy.

State
Customer says they were charged twice.
Questions

Noul: Is the customer requesting a refund?

Choice: Which team should handle the request?

Score: How frustrated is the customer?

Result
Typed, probabilistic decisions.
Application
Route, verify, ask for confirmation, or escalate.

Bounded decisions

Where Jev fits

Agent routing

Choose a tool, subagent, or next workflow branch from known options.

Support triage

Classify a request, score urgency, and identify when review is needed.

Verification

Judge whether a claim, action, or output satisfies a defined condition.

Decision gates

Turn narrow judgments into act, confirm, escalate, or stop policies.

When to use Jev

Start with the mental model: Which one? Where on this scale? Is this true?

  • The answer space can be defined in advance.
  • The same kind of judgment is needed repeatedly.
  • Software will consume the result.
  • Uncertainty should affect what happens next.
  • The task can be decomposed into narrow questions.

When not to use Jev

Use a generative or reasoning model when the primary job is to create, converse, or explore an open-ended answer.

  • Article writing or long-form explanations
  • Open-ended chat
  • Code or creative generation
  • Tasks whose useful answer cannot be bounded in advance
  • Complex open-ended reasoning that cannot be decomposed into focused judgments

Accuracy boundary

Can Jev still make a wrong decision?

Yes. Typed and constrained output does not mean every judgment is correct.

Typed does not mean correct

Jev can choose the wrong valid option, assign an inaccurate score, or return an inaccurate probability while still matching the required answer shape.

Calibration is statistical

Calibration is assessed across groups of predictions. A value of 0.90 is not a guarantee that one particular answer is correct 90% of the time.

Policy belongs in code

Your application decides whether to act automatically, ask for confirmation, escalate to a person, or call another model. Thresholds should reflect the risk.

Vendor reported

How to read “can't hallucinate”

TypeSafe uses “can't hallucinate” and “zero hallucinations” for Jev. A precise interpretation is that Jev's output is constrained to the answer space defined by the developer, so it does not invent an unsupported output type or option. This does not mean every valid decision is correct.

Evidence

Claims and examples are labeled

Product definitions and primitive behavior come from current TypeSafe documentation. TypeSafe's performance and hallucination language remains vendor-reported. JevKit scenarios and values are illustrative unless a reproducible test says otherwise.

OfficialVendor reportedIllustrative

Sources

Try the structure

Turn a workflow into typed questions

Practice the structure of typed decisions with JevKit's generic Decision Schema Builder. It is a learning tool, not an official Jev schema or API payload builder.

Build a Decision SchemaBack to JevKit