TypeSafeのJevを正しく驚く、それってLLMでできませんか？
Japanese article testing whether LLMs can replicate Jev, via a logit shortcut on Gemma and the Mario harness.
This summary is based on how curated lists describe the project; its README was not available to review.
Overview
This Japanese article on Zenn, titled roughly "Being properly surprised by TypeSafe's Jev: can't an LLM do that?", walks through what Jev is and is not. According to the list summaries, it reproduces the shortcut of reading option probabilities from logits instead of generating JSON on a Gemma model, and compares Jev with LLMs on the public Super Mario harness.
It is useful for readers who want a skeptical, hands-on look at how much of Jev's behavior can be recreated with open LLMs.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which Mario action to take (compared across Jev and LLMs)(inferred) | — |
How it works
- Explain Jev's typed-decision approach.
- Reproduce logit-based option scoring on Gemma.
- Compare Jev and LLMs on the Mario harness.
Strengths
- Hands-on reproduction rather than opinion.
Caveats
- Japanese only.
- Known through list summaries.
At a glance
- Type
- Article
- Status
- discussion
Use cases and integrations
How curated lists describe it
Reproduces the JSON-vs-logit shortcut on Gemma and compares Jev with LLMs on the public Mario harness.
AnotiaWang/awesome-decision-modelsJapanese walkthrough of what Jev is and is not.
hellogumbo/awesome-jev
Frequently asked questions
What does the article test?
Whether reading logits from an LLM (Gemma) can replicate Jev-style decisions, and how both do on the Mario harness.
What language is it in?
Japanese.
Where is it published?
On Zenn.
Related projects
Python library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePythonLocal Snake game where code computes facts for each legal move and a local Qwen3.5 picks one in a single Jev-style Choice pass.
AppprototypePythonX thread arguing Jev's speed comes from parallel decoding, so a Jev-like API can sit over any open-weight model.
Discussiondiscussion- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
ProjectbetaGoGo gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
ServicebetaGo