Can I drop these LangGraph routes into production?

No. The digest states it is not production code, thresholds are untuned starting points, and results come from small hand-written scenario sets, not a benchmark of accuracy.

Does the control experiment prove Jev beats chat models?

No. A control experiment on the same questions found Jev comparable to strong structured-output settings on accuracy, with an edge on consistency and speed, within stated limits and run modes.

Can the cases run against a local LM Studio model, or do I need ChatGPT or NIM?

A local model works: the published runs used LM Studio with google/gemma-4-e4b, where only the answer route spent chat-model time, 110 s on average. The ChatGPT subscription provider is called experimental and unofficial, NVIDIA NIM needs NVIDIA_API_KEY, and every live path needs a TypeSafe key; the offline tests run with neither.

Why does only one route spend chat-model tokens?

Code owns the workflow and Jev makes fast typed judgments first; only the answer path spends chat-model tokens while escalate, review, and refuse paths avoid generation cost.

What routing policy turns Jev's answers into a route?

Injection at or above 0.8 gives refuse, urgency at or above 1.5 gives escalate, an unclear intent such as other or confidence below 0.5 gives review, and everything else gives answer.

Which chat-model providers can the pilot run against?

A ChatGPT subscription, NVIDIA NIM, or LM Studio. Set LLM_PROVIDER, or leave it unset to use the first one that is configured. The ChatGPT path is described as experimental and unofficial.

What was the control experiment's finding?

Across 60 messages nothing beat Jev on accuracy (Jev 0.967, the best ChatGPT setting also 0.967), and Jev was more consistent (1.000 against 0.94 to 0.997) and faster (0.6 s against 2.3 s and up).

CommunityJev Models & Reimplementations

Pilot Typesafeai Jev

Pilot wiring Jev into LangGraph routing and Deep Agent guardrails with executed notebooks, showing where code, Jev judgments, and chat-model replies each belong.

See highlights

Open github.com

Stars: 2
Language: Jupyter Notebook
License: MIT
Published by: Community
Author: jyje
Last verified: Oct 2026

jyje/pilot-typesafeai-jev

User repository on GitHub

View on GitHub

👩‍🔬 Pilot of the decision model 'jev' from TypeSafe AI

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

🚀 Pilot project for TypeSafe AI Jev in LangGraph and Deep Agents, on ChatGPT, NVIDIA NIM, or LM Studio. English / 한국어 / 日本語 / 简体中文 / Docs. Found this useful? Please give it a ⭐. It helps others find it. Put TypeSafe AI's Jev into LangGraph and Deep Agents for real, and record what works and what does not. Understand Jev. It reads text and returns typed answers with probabilities (Choice, Score, Noul) instead of generated text, and every answer carries a confidence. See the Jev overview.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Highlights

Separates duties: code owns flow, Jev judges quickly, and a chat model writes replies.
Routes in LangGraph with one Jev call asking intent, urgency, and injection questions.
Guards Deep Agents before start and verifies claims with supported, contradicted, or unrelated verdicts.
Publishes executed notebooks, verification notes, and a structured-output control experiment.

Watch out

Pilot code with untuned thresholds and an experimental subscription provider, not production routing. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

system-one-foundation-models

Native Swift 6 package that evaluates @Generable structs and enums through System One decision models in 15-150 ms, with on-device Laya, self-hosted laya-serve, and TypeSafe Jev cloud backends.