Jev Case Study: Spam Comment Detection (Full Code)

TL;DR: This Jev example runs spam comment detection as a single judgment question on Jev 1.13. A yes above your confidence threshold auto-hides the comment; anything lower goes to the moderation queue. Full request body, example-fixture response, and a copy-paste curl command below.

Scenario

Our engineering blog accepts comments under every post. Every few weeks a spam wave hits — replica-watch ads, casino links, SEO gibberish that names the article topic once and then pivots to a discount site. Keyword filters catch the lazy ones; anything with light obfuscation (“best-deals-now dot club”) sails through, and the team wastes a morning cleaning up.

The fix is deliberately boring: one Jev judgment call per comment, at ingestion time, before the comment is ever shown.

• If the answer is yes and confidence is at least 0.9, hide the comment automatically and log the rationale.
• If the answer is no with high confidence, publish immediately.
• Anything else — low confidence, or an unclear answer — goes to the human moderation queue.

This is exactly the shape of problem Jev is built for: a judgment question (yes / no / unclear) with a typed JSON answer and a confidence value, not a chat reply that has to be parsed out of prose. Jev is TypeSafe AI’s System One judgment model, released September 2026; this example runs it through the OpenRouter OpenAI-compatible endpoint.

One design note: we do not ask the model to write anything. The question is fixed, the answer space is fixed, and the confidence score is what the gate is built on.

Request

{
  "model": "typesafe/jev-1.13",
  "messages": [
    {
      "role": "system",
      "content": "Judgment question. Is this comment spam? Answer yes, no, or unclear. Reply with JSON only: {\"answer\": <yes|no|unclear>, \"confidence\": <0-1>, \"rationale\": <one sentence>}"
    },
    {
      "role": "user",
      "content": "Great article! Also check best-deals-now dot club for 90% off designer bags, fast shipping worldwide!!!"
    }
  ]
}

Response

Example fixture — illustrative output, not a live capture:

{
  "answer": "yes",
  "confidence": 0.97,
  "rationale": "The comment promotes an external discount site, uses urgency and punctuation patterns typical of spam, and has no substantive relation to the article topic."
}

The only fields the gate needs are answer and confidence. The rationale goes straight into the moderation audit log so a reviewer can confirm the auto-hide in one glance.

Reproduce

Set OPENROUTER_API_KEY and run:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-1.13",
    "messages": [
      {
        "role": "system",
        "content": "Judgment question. Is this comment spam? Answer yes, no, or unclear. Reply with JSON only: {\"answer\": <yes|no|unclear>, \"confidence\": <0-1>, \"rationale\": <one sentence>}"
      },
      {
        "role": "user",
        "content": "Great article! Also check best-deals-now dot club for 90% off designer bags, fast shipping worldwide!!!"
      }
    ]
  }'

The JSON body of the response arrives inside the standard OpenAI-compatible chat completion structure; parse the message content as JSON to get the fields above.

Key parameters

Notes

• Responses are example fixtures; verify field names against official docs before relying on them. Official API details live at typesafe.ai — see the System One announcement post.
• Confirm the exact model slug (typesafe/jev-1.13) on the OpenRouter model page before shipping.
• Start the auto-hide threshold conservative (e.g. 0.95) and lower it only after reviewing a week of rationales from the moderation queue. A false positive hides a real reader; a false negative just costs a moderator a click.
• Keep unclear as its own path — do not collapse it into no.
• Applies to Jev 1.13 (released September 2026) via the OpenRouter demonstration channel.