jev-vs-laya
Korean study notes and TypeScript mock-ups of Jev concepts, with a local Laya versus GPT-5 comparison and a calibration toolkit.
This repository is mostly documentation or write-up rather than runnable code.
Overview
jev-vs-laya is a Korean-language study repository about Jev. Because Jev was in early access, it aims to understand the published concepts by turning them into code rather than calling the real API by default. It contains nine study notes covering what Jev is, how state, questions and the three primitives work, how it compares with LLMs (RLHF, RLVR and RLCD), use cases and limits, references, a dissection of the open jevlike implementation, hands-on local training of a decision model on the repo's ticket data, and two notes on Laya, an open-source Jev alternative.
The code models Choice, Score and Noul and confidence as TypeScript types, provides a deterministic mock client, demonstrates confidence-gated routing where ambiguous tickets escalate to humans below a 0.5 threshold, and includes calibration tools for Brier score, reliability buckets and ECE. A real Jev client via the Requesty router is also included. An example and interactive demo compare local Laya with GPT-5 on the same tickets, and the notes report Laya as 26 times faster at no cost. There are 37 tests.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should this support ticket be routed to? | billing · technical · account |
How it works
- Model the primitives as TypeScript types
- Route mock tickets with confidence gates
- Measure calibration with Brier and ECE
- Compare local Laya with GPT-5 on the same tickets
Strengths
- Runs without an API key via mocks
- Calibration measurement utilities
- 37 tests
Caveats
- Study project, not a product
- Notes are in Korean
Getting started
git clone, npm install, npm test, npm run example.At a glance
- Type
- Project
- Status
- research
- Language
- TypeScript
- Runtime
- Node
- Has code
- yes
- Has tests
- yes
- Stars
- 0
Use cases and integrations
How curated lists describe it
Jev(TypeSafe AI System One Model) 스터디 — 타입화된 결정·RLCD·confidence-gated routing을 한국어 노트와 TypeScript 목업으로 정리
wh000wh000/awesome-jev-live
Frequently asked questions
Does it call the real Jev API?
Mostly no; it uses mocks, though a Requesty-routed Jev client is included.
What is Laya?
An open-source Jev alternative that runs locally, compared here with GPT-5.
What language are the notes in?
Korean.
Related projects
Learning project that builds choose-an-answer decision models from scratch and from Qwen, with downloadable 0.8B and 4B releases.
LibraryresearchPythonPreprint survey of Jev-style typed decision models: their research roots, adaptation costs and what current evidence supports.
ModelresearchInteractive web workbench for exploring and benchmarking Jev's Choice, Noul and Score via OpenRouter, with threshold transfer curves.
ProjectprototypeTypeScript- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython 33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonJoke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPython