Jev evaluations

Jev is a dedicated evaluation model served through Vercel AI Gateway. It answers typed questions about supplied state instead of generating conversational text. Use it when an agent needs a probability it can compare with a threshold, record, or pass to the next step.

Naven exposes Jev as a normal x402-paid endpoint:

POST https://agents.naven.network/ai-gateway/v1/evaluations

Each bounded evaluation costs 0.003 USDG. The price is fixed for the request and includes Naven's service fee. Payment settles only after Jev returns a valid evaluation.

Boolean decision

{
  "model": "typesafe-ai/jev",
  "state": "A deployment failed twice after a migration. Rollback succeeded and no customer data was lost.",
  "questions": {
    "escalate": {
      "type": "boolean",
      "instructions": "Should this incident receive immediate human review?",
      "criteria": {
        "true": "Immediate review is warranted.",
        "false": "The workflow can continue automatically."
      }
    }
  }
}

A successful answer contains the probability of true:

{
  "object": "evaluation",
  "model": "typesafe-ai/jev",
  "answers": {
    "escalate": { "type": "boolean", "probability": 0.91 }
  },
  "usage": {
    "inputTokens": 42,
    "outputTokens": 8,
    "totalTokens": 50
  },
  "rounding": {
    "probabilityDecimals": 4,
    "scoreDecimals": 4
  }
}

Treat the probability as model output, not as a factual guarantee. Define and test thresholds for your application. Keep a human in the loop for high-impact or irreversible actions.

Question types

Boolean

Use type: "boolean" for a yes-or-no decision. criteria.true and criteria.false are optional descriptions that make the intended boundary explicit. The answer contains probability, the probability of true.

Choice

Use type: "choice" with 2–16 named criteria:

{
  "type": "choice",
  "instructions": "Route this support request.",
  "criteria": {
    "billing": "Payment, invoice, or refund issue.",
    "technical": "Product behavior or integration issue.",
    "security": "Possible account or asset compromise."
  }
}

The answer contains the selected choice and can include a probability for every choice.

Score

Use type: "score" with 2–10 ordered descriptions:

{
  "type": "score",
  "instructions": "Score the operational risk.",
  "criteria": [
    "No material risk.",
    "Limited and reversible impact.",
    "Material impact requiring review.",
    "Critical or irreversible impact."
  ]
}

The answer contains a numeric score and can include the probability distribution used to calculate it.

Limits

Field	Limit
`model`	`typesafe-ai/jev`; it may be omitted because this is the only evaluation model.
`state`	Nonempty text.
`questions`	1–8 uniquely named questions.
Question ID / choice key	Starts with a letter; then letters, numbers, `_`, or `-`; at most 64 characters.
`instructions` / criterion text	Nonempty; at most 512 characters each.
Combined `state` and `questions`	At most 8,192 UTF-8 bytes after JSON encoding.
Request body	At most 65,536 bytes including JSON framing.

Unknown request fields are rejected. Multiple questions share one evaluation price as long as the request stays within these limits.

Availability

GET /models includes an evaluation object with Jev's ID, price, atomic amount, description, and availability. Jev does not appear in the data array because that array is specifically the Chat Completions catalog.