Jev-Style Typed Decisions Without Generation

The build turns each schema field into a lettered question and reads option logits from a frozen Qwen3-4B, without generating output tokens. It reports the same accuracy as grammar-constrained JSON on closed enums.

View on X time4× faster on short inputs, up to 2.4× on long ones with a shared-prefix cache

Michael Rakutko@m_rakutko𝕏

Jev-style "typed decisions" without training anything. On a frozen Qwen3-4B: turn each schema field into a lettered question and read the option letters' logits. No output tokens. Same accuracy as grammar-constrained JSON on closed enums. 4× faster on short inputs, up to 2.4× on long ones with a shared-prefix cache. Strings and numbers still need generation. Write-up, figures, teaching bench: http

Sep 17, 2026X postsView on X

Strings and numbers still require generation. A shared-prefix cache is used for the reported long-input speedup.

Also filed under Tools & apps

TypeSafe Jev Decision-Model Client
It provides a shared TypeSafe Jev decision-model client for the dsh ecosystem, with strict response validation, keychain-first key resolution, and typed fail-open errors. It is Jev-optional by design.
Jev Reviews Changed Code
The Claude Code plugin uses Jev (TypeSafe AI) to assess changed code at the end of a turn and, when needed, propose a code review using Open Code Review rules.
Alerts on Company Data-Use Changes
TOS Watch sends an email when a company changes what it does with your data. It combines Open Terms Archive history, a Jev classifier, a static site, and a Cloudflare Worker.
OpenAI API-to-Jev Gateway
jev-gateway turns any OpenAI API backend into a gateway compatible with Jev.