Is this a benchmark showing which model is better?

No, and the README says so directly: it is a learning demo, and the two systems do different jobs. The comparison is also lopsided by construction, since Jev always returns the same three fixed Nouls (is_urgent, needs_human, is_financial) whatever the input is, while GPT writes free text against a 75-character prompt.

What does a Noul value of 0.5 actually mean?

That yes and no are similarly likely, not that the model feels moderately strongly. Noul is P(yes) on a zero to one scale with no separate confidence field, so a value near 1 is a strong yes and near 0 a strong no.

Can I try it before I have any API keys?

Yes. The preset example buttons replay saved GPT and Jev answers and make no API calls, and the README notes the deployed examples work without keys. Typing your own input and clicking Run comparison needs both OPENAI_API_KEY and TYPESAFE_API_KEY configured server-side, and the same limit applies there: two custom live runs per visitor per 24 hours.

Where do the API keys live, and can the browser reach them?

The browser posts to /api/arena and a Node route handler calls the OpenAI and TypeSafe SDKs, so both keys stay on the server. The README is explicit that they must not be given a NEXT_PUBLIC_ prefix, which is what keeps them out of the client bundle.

How do I run it locally?

It needs Node.js 20+. Copy .env.example to .env.local and add OPENAI_API_KEY and TYPESAFE_API_KEY, then npm install and npm run dev, and open http://localhost:3000. An optional Python smoke test for Jev needs uv.

Which questions does the Jev side actually ask?

The same three support-ticket Nouls every time: is_urgent (Does this convey urgency?), needs_human (Should a human support agent review this?) and is_financial (Is this about a financial transaction?). There is no separate confidence field, and the Jev side barely responds to what you type.

CommunityJev Demos & Experiments

GPT Vs Jev

A small Next.js demo that sends the same input to GPT and Jev side by side, showing generative text next to three fixed Noul probabilities.

See highlights

Open github.com

Stars: 1
Language: TypeScript
License: MIT
Published by: Community
Author: TanayPadar
Last verified: Oct 2026

TanayPadar/gpt-vs-jev

User repository on GitHub

View on GitHub

Compare GPT generated language with JEV structured Noul decisions on the same input.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Created by Tanay Padar. Open source under MIT. This app sends the same input to two different kinds of model and shows the outputs side by side: JEV (TypeSafe System One) returns structured Noul answers: the probability that the answer is yes. The models are not chained. GPT never writes JEV’s answer, and JEV never writes GPT’s. It is a learning demo, not a claim that one model is better. They do different jobs.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

GPT and Jev are called independently in a server route handler and never chained, so neither model writes the other's answer.
The Noul field is P(yes) on a zero to one scale, with near 0.5 meaning yes and no are similarly likely rather than medium intensity.
Preset example buttons replay saved answers with no API calls, so the page is usable before you have any keys.
Both keys stay server-side in a Node route handler and must not be prefixed with NEXT_PUBLIC_, so the browser never receives them.

Watch out

This is a learning demo and the README states plainly that it is not a claim that one model is better. Jev always answers the same three fixed support-ticket Nouls regardless of what you type, so the Jev side barely responds to input. Custom input is capped at 75 characters and GPT replies at 80 completion tokens. Live comparisons are limited to two per visitor per 24 hours, keyed on IP plus a cookie and held in memory, so a server restart resets the quota. Cost figures are list-price estimates rather than invoices. There is also an optional Python smoke test for Jev that needs uv. 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

laya-server

Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.