The term "System One" is borrowed from:

Daniel Kahneman's Thinking, Fast and Slow — System 1 is fast and automatic thinking (2 × 2); System 2 is slow and deliberate (17 × 24).

A Jev request is made of:

A state plus a set of typed questions — The state is the data the decision is about; each question carries a type that fixes its possible answers in advance.

Jev's three question types are Noul, Choice and:

Score — Noul is yes/no, Choice picks one option from a list, Score rates on ordered levels such as low to critical.

What does a Noul answer contain?

The probability that the answer is yes — For example noul: 0.9 for "is this a refund request?". There is no separate confidence field — the probability already carries the uncertainty.

A Choice has three options and the top one has probability 0.85. Using TypeSafe's formula, its confidence is about:

0.78 — (p_max − 1/n) / (1 − 1/n) = (0.85 − 0.333) / 0.667 ≈ 0.78 — the value in the docs' quickstart response.

Choice confidence for probabilities (0.6, 0.3, 0.1) and for (0.6, 0.2, 0.2) is:

0.40 and 0.40 — Only the top probability enters the formula. The top-to-second ratio (2× versus 3×) is what tells these two apart.

Why does a Score split (0.5, 0, 0.5) get lower confidence than (0, 0.5, 0.5)?

Levels are ordered, so far-apart doubt costs more — Torn between opposite ends gives 0; torn between neighbouring levels gives 0.25. The Choice formula would score both the same.

A Score question with levels 0–3 returns score: 2.2. That number is:

The probability-weighted average level — The score is the expectation over levels, so it can land between them — here between high and critical.

A model is well calibrated when:

Answers stated at 0.8 come true about 80% of the time — The stated probability matches the observed frequency. That is the diagonal on a reliability diagram.

AI-ENGINEERING

intermediate

Jev & System One Models

AI Engineer path · lesson 108 of 124

Meet Jev, TypeSafe's model that never writes text: send one state and a set of typed questions, and every answer comes back at once with a probability. Compute confidence by hand, read a reliability diagram, and set the thresholds that decide when code acts and when a person looks.

The idea

An experienced mail-room sorter. They do not write letters; they glance at each one and drop it in a pigeonhole — and the good ones can tell you which envelopes they were unsure about. Jev is the sorter. The LLM is the person who writes the reply.

How to use

Start on System 1 and race the two models. Then drag the probability sliders and watch the confidence number move by the published formula.

The name comes from Daniel Kahneman's Thinking, Fast and Slow. System 1 is fast and automatic: 2 × 2 is 4 before you have decided to work it out. System 2 is slow and deliberate: 17 × 24 needs steps (it is 408).

Chatbot LLM

Writes its answer one token at a time.

generates text

Reasoning model

Writes out a chain of thought before answering — the closest AI gets to System 2.

generates more text

Jev

Picks answers from options you defined and gives a probability for each. Writes nothing.

System 1

One email, three quick judgment calls

state“I was charged twice this month. Please fix it.”

LLM → JSONtoken 0 / 13

Jev → typed answerswaiting

All three questions go in one request and come back together. Nothing is written token by token, which is a large part of why it is faster and cheaper on this kind of question. (The race speed here is illustrative; TypeSafe quotes 70–500 ms end to end.)

Key insight

Much of the AI inside software is not System 2 work. “Is this a refund request?” is a glance, not an essay — and a glance does not need a model that writes.

Who should make this call?

Sort

Sort each step of a support workflow by what should handle it.

Tap a piece, then tap where it goes — or drag it. Tap the tray to take one back.

Jev (System One)

Drop items here

Generative LLM

Drop items here

Plain code

Drop items here

7 pieces left to place

Test your knowledge

20 questions

Question 1 of 20

Score: 0

How does Jev answer a question?

The short version

An experienced mail-room sorter. They do not write letters; they glance at each envelope and drop it in a pigeonhole — and the good ones can tell you which envelopes they were unsure about. Jev is the sorter. The LLM is the person who writes the reply.

In more depth

Jev & System One Models

Jev is a model from the startup TypeSafe that answers questions without writing any text. You give it a state and a set of typed questions; it returns an answer to each with a probability.

System 1 and System 2

The name comes from Daniel Kahneman's Thinking, Fast and Slow. System 1 is fast and automatic (2 × 2 = 4). System 2 is slow and deliberate (17 × 24 needs steps). A chatbot writes one token at a time; a reasoning model writes out a chain of thought first — the closest AI gets to System 2. But many decisions inside software are quick judgment calls, and that is the work Jev is built for.

The request

State — the data the decision is about. A string, a JSON object or an array of text: e.g. the support email plus the customer's recent charges.

Questions — each with a type that fixes the answer space in advance:

Noul — yes/no. Returns the probability of yes (noul: 0.9).

Choice — one option from a list you define. Returns the choice, a probability per option, and a confidence.

Score — a rating on ordered levels. Returns the expected level (which can fall between levels), a probability per level, and a confidence.

The state is ingested once and every question is evaluated against it independently, in parallel. All answers come back together; nothing is generated token by token, which is a large part of why it is faster and cheaper than an LLM writing JSON. Because outputs are defined in advance, the answer can only ever be one of your options.

Architecture: what is and is not published

TypeSafe describes a new model architecture, a parallel sampler and a new training method, RLCD. It has not published parameter counts, layers or how the sampler works, so the reliable architecture is the interface: one state, many typed questions, a probability for every option. jev-1.13 accepts text only, with a 64k-token request budget, and is priced on input tokens with output free.

Confidence

Confidence summarises how concentrated the probabilities are, from 0 (even spread) to 1 (certain):

Choice: (p_max − 1/n) / (1 − 1/n). Top of three at 0.85 → 0.78.

Score: 1 − (probability-weighted distance from the most likely level) / (the same for an even spread, from the middle level), floored at 0. Doubt between neighbours costs less than doubt between ends.

Noul: no separate field — the probability already carries the uncertainty. |2p − 1| puts it on the same scale.

Calibration and RLCD

A calibrated model's 0.8 answers are right about 80% of the time — a property of many predictions, plotted as a reliability diagram against the diagonal.

RLHF (from human feedback): raters pick the response they prefer. People like confident answers, so a model can learn to sound sure when it is wrong.

RLVR (with verifiable rewards): automatic checks — did the maths come out right, do the tests pass. Rewards the final answer, not knowing how sure to be.

RLCD (for calibrated decisions): the model gives a probability per option and is rewarded when those probabilities turn out to be right.

From probabilities to code

Two thresholds give three paths: act automatically above the top one, do nothing below the bottom one, ask a person in between. Raise the bar as the cost of a mistake rises. Because the numbers are calibrated, the threshold also tells you roughly how often the automatic path will be wrong.

In a workflow

Jev does not replace the LLM. Jev triages the email (System 1), the LLM drafts the reply (System 2), Jev classifies the customer's response. It also works as a guardrail, screening every message into and out of a chatbot for things like jailbreak attempts.

Limits: text only; weak at maths, counting and date comparison (keep those in code); can be steered by instructions hidden in the data; reads instructions literally; may lean toward the first option of a Choice.

The name

After William Stanley Jevons, who observed in 1865 that more efficient steam engines led Britain to burn more coal — the Jevons paradox. Make a judgment call fast and cheap enough and it goes where an LLM never could: every row of a database, every line of a log file.