Does moderation need the full thread?

Usually yes — sarcasm and context read differently with a couple of parent comments. States run up to 100,000 characters, so including context is cheap: it is one request either way.

Can it detect coordinated harassment campaigns?

No — this decides one item at a time. Campaign patterns need aggregation in your platform; typed decisions still feed that pipeline. Store the threat flag per item and analyze across accounts.

Not in this call — Jev is text only. OCR or transcribe text into the state and the same decision shape works.

Use case

Content moderation

Decide allow, review or remove, and flag real threats separately.

The problem

Keyword filters over-block slang and miss doxxing dressed as a joke.

What your code does with the answer

Remove on high-confidence violations. Queue the rest for a moderator.

The two questions moderation actually needs

One choice — allow, review or remove — carries the policy posture; one noul asks separately whether the comment encourages harassment or violence against a person. Splitting them matters: the example comment is rude the way internet comments are rude, while the "show up at his house" clause makes it a removal with a threat flag. One question collapses that distinction; two keep it, with probabilities that let a moderator see contested cases instead of silently mislabeling them.

From probabilities to enforcement

Remove on a high-confidence removal; queue the rest for a moderator with the distribution attached — a 50/50 allow/review split is the item a moderator wants to see next. Keep the threat flag firing even when the action reads allow: harassment signals drive escalation and user-safety paths regardless of where the content lands. Fix thresholds that misfire by fixing the criteria wording, not the number.

Pitfalls specific to moderation

Moderation criteria drift with slang — review the allow/review/remove wording on a regular cycle. Watch for one-shot moderation on contested cases: a 45/55 allow-review split you label silently as "allow" is how review queues get bypassed exactly where they matter. Keep a person on the contested band and log the distribution, so patterns in later misjudgments stay visible.

The request shape

This is the full request the playground sends for this preset — the same JSON your server sends to /v1/systemone with a Bearer key.

curl

curl https://jevai.website/api/v1/systemone \
  -H "Authorization: Bearer $JEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "state": "lol this guy has no idea what he is talking about, someone should dox him and show up at his house",
  "questions": {
    "action": {
      "type": "choice",
      "instructions": "What should happen to this comment?",
      "criteria": {
        "allow": "Acceptable, even if rude",
        "review": "Borderline, needs a moderator",
        "remove": "Clear policy violation"
      }
    },
    "threat": {
      "type": "noul",
      "instructions": "Does the comment encourage harassment or violence against a person?"
    }
  }
}'

FAQ about this use case

Does moderation need the full thread?: Usually yes — sarcasm and context read differently with a couple of parent comments. States run up to 100,000 characters, so including context is cheap: it is one request either way.
Can it detect coordinated harassment campaigns?: No — this decides one item at a time. Campaign patterns need aggregation in your platform; typed decisions still feed that pipeline. Store the threat flag per item and analyze across accounts.
What about images?: Not in this call — Jev is text only. OCR or transcribe text into the state and the same decision shape works.

Run this in the playground