内容
内容审核
用 Jev 将 UGC 判定为放行、复审或拦截。
Choice + Score + Noul
内容审核安全人机协同
这次发给 Jev
点运行时,站点把这份 JSON 发到 POST /api/playground。运行需要登录。
POST /api/playground{
"state": "这家店是诈骗!!!钱被卷走了还在私信里威胁我。所有人避开 @shopXYZ —— 立刻举报。",
"questions": {
"action": {
"type": "choice",
"instructions": "What moderation action should the platform take?",
"criteria": {
"allow": "Safe to publish",
"review": "Queue for human review",
"block": "Block or remove"
}
},
"confidence": {
"type": "score",
"instructions": "Confidence in the recommended moderation action",
"criteria": [
"Low — ambiguous or context missing",
"Medium — reasonable but not definitive",
"High — clear policy alignment"
]
},
"human_review_required": {
"type": "noul",
"instructions": "Should a human moderator confirm before enforcing the action?",
"criteria": {
"true": "Safety, legal, or reputational risk needs human eyes",
"false": "Automated enforcement is acceptable"
}
}
}
}