AI moderation where you set the line.
Every message gets a number, not an opinion.
jevmod scores each message for spam, scams, harassment and six more, and hands your code a probability between 0 and 1.
Try it on your own messageA message arrives.
It is read the moment it is posted, like every other message in the channel.
It is scored, nine numbers.
One probability per category, between 0 and 1. No sentence comes back with it.
You set the line.
One line per category. A score above it crosses it.
What crossed it is flagged, and you decide what happens next.
Flag is the default. Delete and a ten minute timeout are yours to turn on.
Nine categories. Seven are on by default.
Spam, scams, harassment, adult content, self harm, doxxing and comments about minors are on. Off topic and written by a model are off until you turn them on. Each category you have on gets its own probability from 0 to 1.
You can also add up to five rules in your own words: No politics. News about the game is fine.
It flags what reads as written by an AI assistant.
It asks whether a message was written by a language model rather than typed by a person, and it reads the signs an assistant leaves: generic politeness, list-like phrasing, a tidy summary, an explaining register aimed at nobody in particular, no slang. A person still reads it before anything happens.
- Off by default. You turn it on.
- It only ever flags. Delete and timeout cannot be attached to it.
- No false positives on 90 real chat messages. It can also fire on a person who writes like an encyclopedia, so a flag asks a moderator to read, never decides.
You set the line. It flags, and nothing else, until you say otherwise.
Drag the line. Everything to the right of it crosses it.
Flag is the default for all seven. A fresh install deletes nothing.
A score is not fixed. Measured, the same message moves by about a tenth depending on what the model reads alongside it. The line is a policy you set, not a fact about the message.
It scores.0.00
You set the line.0.75
You decide what happens at it.
Flag is the default for the seven categories that start on, so a fresh install deletes nothing. Deleting and timing out are steps you turn on yourself.
What it will never do.
Ban anyone
It has no ban permission and no ban command. The most it can do is delete a message, or time somebody out, ten minutes by default, if you give it that permission.
Judge your moderators
Roles you mark as trusted are never judged, and neither is anyone who can manage messages.
Act when it is unsure
If the model does not answer, nothing happens to that batch. Nothing is removed on a guess.
Make the final call for you
It scores, and it points at what crossed your line. It acts only in the ways you turned on, and anything past that, a ban included, is a moderator's call.
Trained on 905,197 messages from more than 30 sources.
A 140 million parameter model, 144 MB, running on jevmod's own server.
Measured on 2,531 public messages.
At the line it ships with, it catches 78 of every 100 messages that humans labelled as harassment. It misses the other 22.
Harassment caught at the shipped threshold, against human labels.
Per message, inside a request carrying fifty of them. The batch itself takes 801 ms at the median and 1,110 ms at the 95th percentile, over 20 runs.
Words of prose returned, and therefore nothing to hallucinate. Every answer is a number between 0 and 1, per category, per message.
Latency measured over 20 batches on 21 September 2026, steady state rather than the first call of a cold process, from one machine on one network.
One message is rarely the problem.
A raid or a spam wave is a pattern across many messages. That is what we are building next, and it keeps the same rule: jevmod points at the pattern, and a moderator decides what to do about it.
Illustration of work in progress. The queue ranks three recorded example messages by the score they already have.
14 days to try it. Then $3.99 a month, per server.
Subscription
One server, one subscription.
14 days of the whole thing, every category and every rule. A card is required to start it, and a server gets one.
$0Per server, per month. 50,000 judged messages, the log channel and the dashboard. Cancel any time in Stripe's portal.
Or skip all of it and call the API.
One HTTP POST, or an MCP server your coding agent can call. MIT, and the same decision behind every one of them.
Try it on a message of your own.
Type anything, or pick an example. It is scored by the same service a server runs on, and comes back as numbers.
What you send is kept for 90 days to stop abuse. Privacy notice.