Esta guía está disponible actualmente en inglés y chino.

Workflows / JevModel

Use Jev for review queues

Triage posts for review without treating a probability as a final verdict.

Última actualización:

Ask about a policy, not a feeling

State the applicable policy in the question and define what a positive Noul result means. For complex policies, split categories into separate questions rather than asking for a vague overall judgment.

Try a specific policy question

After email verification, create a Console API key and POST the JSON below to https://jevmodel.app/api/v1/systemone with Authorization: Bearer YOUR_JEVMODEL_KEY and Content-Type: application/json. This example asks whether a message needs review under one narrow harassment policy. It does not ask the model to ban a user or decide the final penalty. Adapt the policy to your community and keep keys out of browser code.

{
  "state": "You are useless and I will keep messaging you until you leave this community.",
  "questions": {
    "needs_review": {
      "type": "noul",
      "instructions": "Does this message need human review for targeted harassment under the policy described below?",
      "criteria": {
        "true": "A targeted personal insult, threat or repeated unwanted contact directed at a person.",
        "false": "Good-faith disagreement, product criticism or a benign message without targeted harassment."
      }
    }
  }
}

Interpret Noul and test the boundary

Read data.answers.needs_review.noul from the successful response. An illustrative value of 0.91 is the model’s probability for the true description, not a measured 91% accuracy rate; Noul has no separate confidence field. The same words may be quoted for reporting abuse or used as an insult. Include that context in state, test both cases, and do not treat a low value as proof that a post is safe. Choose a review threshold from your own labeled examples.

Audit errors before automating

Use Console → Batch to run the same question over representative CSV, TXT or JSONL inputs, then export the answers. Compare them with independent reviewer labels outside the site and count missed violations and benign posts sent to review separately. Include quotations, reclaimed language, supported languages and ambiguous cases. Batch provides answers, not a built-in accuracy or compliance report. A moderation flow can return review or continue outcomes; your application controls publishing and enforcement.

Keep human control

Use the result to prioritize review. Document thresholds, sampling, appeals, and error rates for each category. Where a decision can materially affect a person, review the evidence and context before enforcement.

Reuse the human-review pattern

The linked GitHub human-review example demonstrates the same Noul review-queue pattern using a billing complaint rather than a community post. Change state and the true/false criteria to match your moderation policy. Run from the repository root; add --live after setting JEVMODEL_API_KEY only for one real request. Inspect data.answers.escalate.noul, a 0–1 probability with no separate confidence field. The default result is illustrative. Evaluate thresholds using labeled cases; the example performs no refunds, notifications or enforcement.

Batch text classification with the JevModel APIAdd a decision gate before tool useJevModel API: request and responseZendesk AI Ticket Classification & Routing: Built-in Intelligent Triage vs a Custom ClassifierSentiment Analysis API in 2026: Cloud APIs vs LLM Prompts vs a Typed Decision ModelCustomer Feedback Classification: One Theme, a Priority Score, and Why Sentiment Is Not EnoughSpam Detection API for Forms and Comments: Turnstile, Akismet, or a Jev Decision Layer
node quickstart/node.mjs human-review
# Or: python quickstart/python.py human-review

JevModel es independiente y no está afiliado a TypeSafe AI.