Jev as a Classifier

/ 01

Short Answer

Jev returns typed decisions rather than generated prose. That makes it useful as a classifier: ask which category an item belongs to, how severe it is, or whether it needs attention, then consume the JSON in code. Choice, Score and Noul provide different answer shapes with probabilities. Current Jev pricing charges input tokens while output tokens are free.

/ 02

Why Jev Fits Classification

Choice returns a named option, a probability distribution and confidence; Score returns a weighted rating over ordered levels; Noul returns one yes/no probability. Your application can read these typed answers without parsing generated prose. At the currently listed $0.042 per million input tokens and free output tokens, short classification requests can cost a fraction of a cent. Check the current OpenRouter model rate before budgeting.

/ 03

Three Primitives, Three Classification Shapes

These question shapes cover categorizing, prioritizing and filtering a triage queue. Score is a weighted level value and may be fractional; it is not inherently a 0–1 scale. Read the Noul guide and Jev JSON output reference for the precise answer shapes.

/ 04

Worked Example — Support Tickets

Include only the fields needed for these decisions. This is a copyable request shape, not a live run or a claim about the resulting probabilities.

{
  "state": { "tickets": [
    { "id": 4412, "subject": "Can't log in after password reset", "plan": "pro", "age_days": 3, "replies": 0 },
    { "id": 4413, "subject": "Invoice shows wrong VAT number", "plan": "pro", "age_days": 1, "replies": 2 }
  ] },
  "model": "jev-latest",
  "questions": {
    "ticket_4412_category": {
      "type": "choice", "instructions": "Which category does ticket 4412 belong to?",
      "criteria": { "account_access": "Login or authentication", "billing": "Invoices or payments", "bug": "Product malfunction", "how_to": "Usage question", "other": "None of these" }
    },
    "ticket_4412_urgency": {
      "type": "score", "instructions": "How urgent is ticket 4412?",
      "criteria": ["Routine", "Prompt attention", "Immediate escalation"]
    },
    "ticket_4413_docs": {
      "type": "noul", "instructions": "Can ticket 4413 be fully resolved by linking the VAT documentation without a person reviewing it?",
      "criteria": { "true": "Documentation fully resolves the issue", "false": "Account-specific correction or investigation is needed" }
    }
  }
}

Each answer returns under the question id you supplied. One playground run is one submission, even when it contains several questions; see Pricing for run allowances and the Jev API guide for direct integration.

/ 05

Worked Example — Alerts and Logs

Use Choice for a named fault class such as timeout, authentication, rate limit, data or unknown. Use Score with ordered severity descriptions, then Noul to estimate whether an alert is actionable. A low or ambiguous probability should go to review according to your own thresholds; Jev returns a judgment, not a verified fact. See Jev’s known limitations.

/ 06

Worked Example — Incoming Feedback

Choice can assign a feedback topic, Score can rate its expected benefit against explicit ordered descriptions, and Noul can ask whether it duplicates an existing request. Compare the model’s answer with your own records before merging or discarding feedback.

/ 07

Keeping the State Clean

Send only fields relevant to the question. TypeSafe documents that irrelevant details in large state can reduce accuracy; prefilter in code, or ask a separate Noul relevance question when needed. Keep counting and arithmetic in code. Read Jev’s known limitations before using it on a production queue.

/ 08

What Classification Is Not a Good Fit For

Do exact counting and arithmetic in code. Very large state with unrelated details also weakens the judgment. Do not treat model output as the final decision in matters with significant effects on people; validate it and keep human review where appropriate. See the service terms.

/ 09

Cost at Queue Scale

Illustration only: 476 input tokens per request × $0.042 per million input tokens; output tokens are currently free. Actual cost changes with input length, model rate, gateway charges and infrastructure. Check Pricing for VerdictKit plans and the current model price.

/ 10

Try a Classification

Build a typed question in the Jev playground or see the Jev AI examples. Inspect the response fields before parsing them in code.

/ FAQ

Frequently asked questions