# Jev cannot write. We gave it a job.

Sent September 22, 2026

Our LLM Selector now uses a model that cannot write a reply. We gave it nineteen questions instead. TypeSafe released Jev on September 15. It reads text and returns choices, scores or yes-or-no probabilities, within the answer space you define. That makes it useful for the small decisions inside an AI workflow, such as routing a ticket or checking a stated criterion. It does not replace a model that has to write the reply.

[Read the Jev breakdown →](/blog/posts/what-is-jev)

## Why the model is getting attention

- [One state, many typed decisions](https://benchlm.ai/blog/posts/what-is-jev): Jev evaluates the questions in a request in parallel. Each answer is a choice, score or yes-or-no probability that code can use directly, rather than prose that has to be recovered from a chat response.
- [The price makes the design worth testing](https://openrouter.ai/typesafe/jev-1.13): OpenRouter listed Jev 1.13 at $0.042 per million input tokens and $0 output on September 22. Questions and their criteria still add input tokens, so fan-out saves repeated state, not the cost of every question.
- [The selector keeps the ranking outside the model](https://benchlm.ai/tools/llm-selector): A description fills the same form a visitor can complete by hand. The deterministic selector code then ranks candidates from public evidence. Jev never sees a score or decides the shortlist.

## The limit

A typed answer can still be wrong. We have not measured Jev’s calibration or speed on our selector traffic. A confidence value does not establish the accuracy of a job classification, and consequential actions still need permissions, confirmation and limits in code.

## The boundary that matters

- The option list is part of the contract.: Jev cannot select an option outside the list you provide. That bounds the output, but a badly defined list can still produce a bad decision.
- Writing remains a separate job.: Jev cannot compose a summary, reply, patch or explanation. Pair it with a language model when a workflow needs both a bounded decision and generated text.

## Analysis worth opening

- [Jev cannot write. That is the feature and the limit.](https://benchlm.ai/blog/posts/what-is-jev): What TypeSafe documents, what the selector sends, how external tests should be read, and what still has to be measured before treating Jev as a reliable decision layer.
- [A valid JSON response can still be wrong](https://benchlm.ai/blog/posts/test-structured-output): Our structured-output evidence run shows why a constrained response shape does not settle semantic accuracy.

## This issue in numbers

- Fixed questions in the selector’s standard request: 19
- Choice and yes-or-no readings in that request: 14 + 5
- Catalog jobs available to the primary job choice: 31 + unknown
- Jev 1.13 input list price on September 22; output listed at $0: $0.042 / M

Archive copy reflects the rankings, prices, and availability stated when this issue was sent. Current pages may show newer evidence.

Canonical page: https://benchlm.ai/newsletter/issues/2026-09-22-jev