MODELS · DECISIONS · AUTOMATION

Jev does not want to chat.
It wants to decide.

TypeSafe presents Jev as the first “System One Model”: it receives unstructured state and questions whose answers are defined in advance, then returns typed values, probabilities, and confidence. It gives up text generation to occupy a smaller, concrete place inside software: the hard if-statements we currently solve with prompts, parsers, and hope.

PUBLISHEDSEP 25 2026

READING~12 min

LEVELTECHNICAL · GUIDED

STATUSEARLY ACCESS

01 / THE INTERFACE

State in.
Decisions out.

The important difference is not cleaner JSON. It is that the output space exists before inference.

STATE

Ticket, alert, document, trace, or product state

QUESTIONS

Choice · Score · Noul

RESULT

Typed values + probabilities + confidence

A chat model can also return JSON, especially with structured outputs. But it still generates a sequence: it emits tokens, may need validation or retries, and confidence is not necessarily part of its contract. Jev proposes a different abstraction. The unit of work is not a textual response but a typed evaluation against shared state.

That also narrows its scope. It does not write an email, explain a diagnosis, generate code, or invent a new option. If the valid answer is missing from the schema, the workflow design—not the model—is incomplete.

02 / THREE PRIMITIVES

Atomic questions.
Composition in code.

Each question should be answerable as a fast, well-scoped judgment. The complete policy lives outside the model.

01 · CHOICE

Choose one option

Returns the selected option, its probability distribution, and a confidence measure.

02 · SCORE

Score against a rubric

Returns a value on a defined scale, plus probabilities and confidence.

03 · NOUL

Evaluate a statement

Returns a number from 0 to 1: how likely the statement is to be true.

STATE

“The customer reports two identical charges. The purchase was today and the second transaction is still pending.”

QUESTIONS

duplicate_likely: Noul
urgency: Score(1..5)
route: Choice([wait, support, fraud])

CODE

if confidence < 0.80:
  send_to_human()
elif route == "fraud":
  freeze_and_review()

The final decision remains deterministic and auditable: the model estimates; the program combines thresholds, rules, permissions, and actions.

03 / WHY PARALLELISM MATTERS

There is no sentence
waiting for the previous one.

TypeSafe says all questions are evaluated at once and independently against the same state.

AUTOREGRESSIVE LLM

t₁ → t₂ → t₃ → … → tₙ

Each token depends on previous tokens. More output text means more sequential steps, more latency, and more surface area for format drift.

JEV · SYSTEM ONE

q₁ ∥ q₂ ∥ q₃ ∥ … ∥ qₙ

Decisions share state but not a text chain. Adding questions should barely affect latency, according to the vendor documentation.

The promise is attractive for workloads with many small decisions: moderating a hundred attributes, scoring fraud signals, routing events, or checking an agent trace. TypeSafe publishes end-to-end latency of 70–500 ms and input pricing of USD 0.042 per million tokens, with output too cheap to meter.

But those numbers are vendor claims and depend on the workload. The published evals use four in-house workflows and answers from large models as reference labels, not independent business ground truth. They are useful evidence for experimentation, not a universal guarantee.

04 / WHERE IT FITS

Beside the LLM.
Not in its place.

Specialization is valuable when the boundary between deciding and generating is explicit.

Routing

Choose which queue, agent, or model should receive a case.

Triage

Prioritize alerts, tickets, claims, or documents by risk and urgency.

Guardrails

Evaluate whether an action or response crosses a policy before executing it.

Closed extraction

Map free-form text to states and categories the program already knows.

INPUT→JEV→POLICY IN CODE→LLM / TOOL / HUMAN

Jev can decide who should act; an LLM can draft; a tool can execute; a person can approve. Separating those roles reduces freedom where reliability matters without giving up generation where it is needed.

05 / WHAT WE KNOW AND WHAT WE DO NOT

A clear interface.
A still-opaque box.

A well-defined product contract should not be confused with a publicly demonstrated architecture.

PUBLISHED

Unstructured state and typed questions as input.
Choice, Score, and Noul as output shapes.
Probabilities, confidence, and parallel evaluation.
A training method called RLCD, aimed at calibrated decisions.

NOT PUBLISHED IN DETAIL

Complete architecture, size, and training data.
A reproducible RLCD recipe and ablations.
Broad independent validation of calibration and benchmarks.
Behavior outside evaluated domains.

HONEST READING“Cannot hallucinate” means it cannot invent a structure outside the schema. It does not mean the in-schema decision is correct. A valid value can still be perfectly wrong; that is why calibration, thresholds, evals, and human fallback matter.

THE BET

Less language.
More software.

Jev’s most valuable idea may not be a benchmark but a boundary: use generative models to produce and decision models to judge; keep policy, permissions, and action in code. If calibration survives beyond the demos, that split could make AI more composable. Until then, Jev is a compelling technical hypothesis that deserves your own evals, not faith.

06 / SOURCES

Announcement, contract
and evidence.

The primary source describes the product; the documentation defines the interface; the evals show the methodology and its limits.

Editorial synthesis based on public information available on September 25, 2026. Performance and cost metrics are attributed to TypeSafe because they do not yet constitute independent validation.