Overview & result
jevmod is moderation for communities and apps. Each message gets a probability for spam, scam, harassment, NSFW, off-topic, self-harm, doxxing and for rules written in plain English; you set the thresholds and actions, and every decision is logged. It is flag-only by default and fails open if Jev is unreachable. The author reports about $0.04 per 1,000 messages with all categories on.
From the original post
Moderation for communities and apps, powered by Jev (TypeSafe): probabilities per category, thresholds you own. Discord/Telegram/Reddit bots, CLI, Python, npm, HTTP API, MCP.
How Jev fits in the loop
- A pre-filter keeps trusted authors, very short messages without links and repeats of already judged text away from Jev.
- Messages are batched for 2 seconds per community and sent in one Jev request.
- Jev returns a probability per category and per custom rule.
- Your policy maps probabilities to actions, and every decision is written to an audit log.
How to reproduce
- Inspect the original showcase and source material at https://github.com/ohernandezdev/jevmod
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
The author documents a bug their red team caught: batching messages as a list leaked probabilities between neighbours, so batches are now keyed by position and tested in CI.
Limitations
- Flag-only by default; nothing is deleted until you turn that on.
- It fails open: if Jev is unreachable, messages are left alone.
- The cost figure of about $0.04 per 1,000 messages is the author's.