Does the model return the verdict, or do you compute it?

You compute it. The README states the verdict is produced in code via policy thresholds, with route() in lib/moderation.ts. The model only returns seven hazard probabilities and a severity score from one systemOne call, so what gets blocked, flagged or allowed is yours to change — and easy to argue about in review.

What happens if I never set an API key?

It runs in mock mode: a keyword heuristic returns a verdict and the response is flagged mock: true. That is handy for getting the UI up, but a keyword heuristic passing can read like a working moderation system. Set TYPESAFE_API_KEY in .env.local, copied from .env.example, to get live calls against jev-latest.

Can I deploy this to production as written?

The README gives you a Vercel path — vercel --prod with TYPESAFE_API_KEY in the env vars — but two stated limits are not production-grade. Text is capped at 1000 characters, and the 10 requests per minute per IP is kept in memory, which the README notes means Redis for multi-instance use. The runtime and server-side key handling are fine; the rate limiter is the piece to replace.

Are the shipped thresholds any good as-is?

The author's own answer is no. The README says thresholds should be fit on labeled traffic, not taken as universal. The seven hazards, the strict and permissive policies and the route() thresholds all live in lib/moderation.ts, and nothing in the README reports how the model scores on real traffic, so treat the defaults as placeholders you have to calibrate before you believe them.

How do I run the moderation guard locally?

bun install, copy .env.example to .env.local and add TYPESAFE_API_KEY (optional, without it you get mock mode), then bun dev and open http://localhost:3000.

What request shape does POST /api/moderate take?

A JSON body with text and a policy of either strict or permissive; the policy selects which thresholds route() applies. Text is capped at 1000 characters and requests at 10 per minute per IP, held in memory.

Where are the hazard checks and thresholds defined?

They live in lib/moderation.ts: HAZARDS, the strict and permissive POLICIES, and route(), which turns the returned probabilities into a verdict. The README says to fit thresholds on labeled traffic rather than use the shipped numbers as universal.

CommunityJev Demos & Experiments

Guard Jev

A Bun and Next.js moderation demo that screens text with one systemOne call — seven hazard Nouls plus a severity Score in parallel — and computes the verdict in code from policy thresholds.

See highlights

Open github.com

Stars: 1
Language: TypeScript
Published by: Community
Author: NorbertBodziony
Last verified: Oct 2026

NorbertBodziony/guard-jev

User repository on GitHub

View on GitHub

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Text moderation demo: one TypeSafe systemOne call screens 7 Noul hazards + 1 severity Score in parallel. Verdict computed in code via policy thresholds. POST /api/moderate { "text": "...", "policy": "strict" | "permissive" }. Max 1000 chars, 10 req/min per IP (in-memory; use Redis for multi-instance prod). Without TYPESAFEAPIKEY returns keyword-heuristic mock with mock: true. With key: live Jev (jev-latest) via @typesafe-ai/sdk, server-side only.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Moderation

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

All eight questions go out in a single systemOne call: seven Noul hazard checks plus one severity Score, evaluated in parallel.
The verdict is not the model's — route() applies policy thresholds from lib/moderation.ts to the returned probabilities.
Two policy levels, strict and permissive, are selected per request in the POST /api/moderate body.
It runs keyless: without TYPESAFE_API_KEY it falls back to a keyword heuristic and flags the response mock: true.

Watch out

The README titles this a demo and the stated limits back that up. Text is capped at 1000 characters and requests at 10 per minute per IP, held in memory — the README itself notes you need Redis once you run more than one instance, so that limit resets on every deploy and is not shared between instances. The individual hazards are not enumerated in the README; they live in HAZARDS inside lib/moderation.ts alongside the strict and permissive POLICIES and the route() function that turns probabilities into a verdict. On tuning the author is direct: thresholds should be fit on labeled traffic, not taken as universal, so the shipped numbers are placeholders for your data. Without a TYPESAFE_API_KEY the app answers from a keyword heuristic and marks the response mock: true, which is convenient for setup but tells you nothing about model behaviour, and mock mode will look like it works. With a key, calls go to jev-latest server-side through @typesafe-ai/sdk, so the key never reaches the browser. 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

AskJev

Claude-driven browser automation where plain-English chat drives Brave or Chrome, Jev makes on-page decisions, and a guard freezes irreversible clicks.