Try it with your own rules
Start from the preset below, adapt the questions, and inspect the typed answers. Run it live in the playground — it is the same request shape your application will send.
What Jev returns
Illustrative output for the preset below — run it live to get real values for your input.
Just got banned from the beta for no reason. The dev team are incompetent cowards who cannot take feedback. I posted a full teardown at my blog, and I am sharing the cracked build in my Telegram group.
What should moderation do with this post?
Does this post violate the community guidelines?
P(yes) = 0.9 — a link to pirated builds, not the harsh criticism, drives the call
How severe is the violation, if any?
From one example to a reusable workflow
Write the guidelines as option definitions
Each action is a choice option with a one-line definition: "Remove: piracy links, ban evasion, spam." Editing that sentence is how you change policy.
Judge the violation separately from the tone
Harsh criticism is allowed; a piracy link is not. Ask the violation question on its own so a heated tone does not push a legitimate complaint into the remove bucket.
Review the middle, not the ends
Auto-handle the confident allow and remove decisions, and route anything between your two thresholds to a person ordered by severity.
Keep the criteria separate
| Check | What it measures | How to use it |
|---|---|---|
| Action · Choice | What should happen to this post? | Option definitions are the moderation policy — keep them short, concrete and mutually exclusive. |
| Violation · Yes/No | Does it break a stated rule? | Threshold by cost: a false remove angers a member, a false allow exposes the community. |
| Severity · Score | How bad is it? | Sort the review queue by severity so the worst items are seen first. |
The problem
Keyword lists miss context, and sending every post to a general chat model is slow and hard to threshold. Moderation needs a fast, bounded decision with a probability you can set a cutoff on.
Keep a human on the middle
Jev is a judgment, not a verdict. Use it to clear the obvious volume and to prioritise the queue; a person should still own bans, appeals and anything the model is unsure about.
Cheap enough to run on every post
A short post with three questions is a few hundred input tokens, so it is practical to check every submission rather than a sample.
FAQ
Can Jev moderate images or video?
Images: yes, experimentally — Jev 1.14 accepts PNG/JPEG input on the preview model, so you can attach a screenshot or frame and ask typed questions about it. Video and audio: no — describe or transcribe them first.
Will it catch sarcasm or coded language?
It reads the whole post in context, but it is probabilistic. Treat low-confidence results as review cases and test it on your own labelled examples before trusting a threshold.
Should the result be shown to users?
Not as a reason. Jev returns a decision and probabilities, not an explanation; write the user-facing message yourself from the action it chose.