Jev — a decision model

Authentication

• Authorizationheaderrequired

  Your Ailoot key in the format Bearer sk-….

Request body

• modelstringrequired

  jev-latest.

• statestring | object | arrayrequired

  What to evaluate: a text string or JSON containing a customer inquiry, a database record, a conversation, or your application's current state. An object with clear field names is usually easier to work with than a single string.

• questionsobjectrequired

  The questions. You choose the keys, and answers are returned under the same keys. The key does not affect the answer; the question itself is specified in instructions.

• questions.*.typestringrequired

  noul, choice or score — see below.

• questions.*.instructionsstring | object | arrayrequired

  The question itself. You can pass a long question with supporting data as an object: put the question in one field and the data in others, and refer to fields by name in backticks.

• questions.*.criteriaobject | array

  Answer options, defined differently for each type: optional descriptions of “yes” and “no” for noul, a dictionary of options for choice, and an ordered list of levels for score.

You can combine question types in a single request.

noul — yes or no

The field criteria is optional: you can use it to describe what “yes” (true) and “no” (false) mean. The answer noul — a number from 0 (no) to 1 (yes).

"is_urgent": {
  "type": "noul",
  "instructions": "Does this convey urgency?",
  "criteria": {
    "true": "Explicitly time-sensitive",
    "false": "No urgency expressed"
  }
}

"is_urgent": {
  "type": "noul",
  "noul": 0.95
}

choice — selecting an option

criteria — an “option → description” dictionary; if no description is needed, use null. Up to 255 options. In the response, choice — the most likely option, probabilities — the probability of each option (summing to 1), confidence — confidence from 0 to 1.

"department": {
  "type": "choice",
  "instructions": "Which team should handle this?",
  "criteria": {
    "billing": "Payments, invoicing, refunds",
    "technical": "Bugs, outages, integrations",
    "sales": "Pricing, upgrades, new customers"
  }
}

"department": {
  "type": "choice",
  "choice": "billing",
  "probabilities": { "billing": 0.88, "technical": 0.12, "sales": 0.0 },
  "confidence": 0.81
}

score — rating on a scale

criteria — levels ordered from lowest to highest, with no more than 10 (TypeSafe recommends at least two). In the response, score — a value weighted by the probabilities of the levels (it can fall between levels), legend — the level number and its description, probabilities and confidence. According to TypeSafe, score is suitable for comparison against a threshold, but not for deriving an exact number between levels.

"frustration": {
  "type": "score",
  "instructions": "How frustrated is the customer?",
  "criteria": ["Calm", "Frustrated", "Very angry"]
}

"frustration": {
  "type": "score",
  "score": 1.05,
  "legend": { "0": "Calm", "1": "Frustrated", "2": "Very angry" },
  "probabilities": { "0": 0.0, "1": 0.95, "2": 0.05 },
  "confidence": 0.92
}

• modelstring

  The model version that answered. The jev-latest alias moves to newer versions over time. If you have calibrated confidence thresholds for a specific version, monitor this field.

• answersobject

  One answer per question, under the same keys. Every answer has type, and choice and score also have confidence.

• usageobject

  input_tokens and output_tokens. Only input tokens are charged.

Submit all questions about the same state in one request: they are evaluated in parallel, and state is sent only once. According to TypeSafe, in their example with 13 questions, one combined request cost less by a factor of 12.2 and was 10 times faster than 13 separate requests sent sequentially, with the same answers.

Break a complex judgment into simple questions and combine the answers in code. For example, instead of “evaluate this application,” ask separately about the market, feasibility, and differences from competitors. You can then adjust the weights in code rather than in the question text.

{
  "model": "jev-latest",
  "state": "Help! My payouts have been failing for 3 days.",
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Payments, invoicing, refunds",
        "technical": "Bugs, outages, integrations",
        "sales": "Pricing, upgrades, new customers"
      }
    },
    "frustration": {
      "type": "score",
      "instructions": "How frustrated is the customer?",
      "criteria": ["Calm", "Frustrated", "Very angry"]
    },
    "is_urgent": {
      "type": "noul",
      "instructions": "Does this convey urgency?"
    }
  }
}

• Routing. Send an inquiry to the right department or a request to the right handler: simple code, a large model, or a human.
• Sorting and prioritization. Rate urgency, sentiment, and risk on a scale, with thresholds defined in code.
• Moderation and checks. Check whether a message violates the rules, a source supports a quotation, or a document contains an answer to a question before passing the text on.
• Scoring and ranking. Evaluate documents, candidates, or retrieved passages against your own rubric, and decide in code what moves forward.
• Replacing fragile “return JSON” prompts. A call that returns typed values by design, rather than text you need to parse.

• Chat, replies to users, and text or code generation. Jev does not generate text; for that, use text models.
• A model for Claude Code, Cursor, and other agents. You cannot replace an agent's model with jev-latest: the agent needs a model that generates text and calls tools. You can call Jev from code the agent writes for you.
• Arithmetic, counting, and date comparisons. According to TypeSafe, these operations are more reliable in code. Ask the model only for the semantic part—for example, extracting date components using choice.
• Images, audio, and video. The model accepts only text; convert these inputs to text beforehand.

Only input tokens are charged, as reported in the usage.input_tokens field of the response: $0.084 per 1 million. Output tokens are free.

The example request on this page (296 input tokens) costs less than $0.0001.

Before the request, Ailoot places a reserve on your balance and, after the response, charges you once based on actual usage — how it works.

Model limits are provided by TypeSafe. Your key and balance limits apply as they do for any Ailoot model: Errors and limits.

Errors use the standard Ailoot API format: a code, a clear message, and a Request ID. The most common errors for this method are:

For the full list of codes, see Errors and limits.