State

What state is

State is the content you ask a System One model to evaluate. It might be a support message, a passage of text, or the current state of your application. You pass it in the state field of an API request, alongside the questions you want answered.

Each request evaluates one state against one or more questions. All questions see the same state and are evaluated independently. You can mix Choice, Score, and Noul questions in one request.

Three shapes

String

The simplest state is a plain string:

state = "My card was charged twice."

Use it when the use case is simple and needs only one piece of text.

Object

When a decision requires comparing several parts, put the related information together in an object, with a descriptive name for each part:

{
  "ticket": {
    "subject": "Duplicate charge",
    "messages": [
      {"from": "customer", "text": "I was charged twice for order A-104. Please refund the duplicate."},
      {"from": "support", "text": "We are checking the charges."}
    ]
  },
  "order": {
    "id": "A-104",
    "charges": [
      {"amount_usd": 49, "status": "captured"},
      {"amount_usd": 49, "status": "captured"}
    ]
  },
  "refund_policy": "Duplicate charges are eligible for a refund."
}

This is one state, even though it contains a conversation, an order, and a policy. The guidance is to use an object for most requests, so each part has a descriptive name and its relationships stay clear.

Array

For a sequence of messages or records:

state = ["Hi", "My customer number is TS1337.", "My card was charged twice."]

Format	Useful for	Example
String	A message, article, or passage	"My card was charged twice."
Object	Named fields, related records, application state	see the JSON above
Array	A sequence of messages or records	see the array above

Separate content from questions

This is one of the most important mental models when using Jev:

• State holds the content and supporting facts. The refund request, the order records, the refund policy.
• Questions define the judgements to make. “Was a refund requested?”, “Does the policy support a refund?”

Do not put judgement logic into the state. State is the material you would present to a panel of experts before asking them to make a judgement.

Language support

Jev accepts text only. State must be a string, JSON object, or array of text values.

• Images, audio, and video are not supported.
• Non-text input must be pre-processed into text or structured fields before being sent as state.
• Jev’s primary training language is English. Other languages, including CJK scripts, are accepted but currently have lower accuracy.

The last point matters if your content is not English: validate accuracy against your own data before putting Jev on a critical path. For high-stakes decisions, consider supplementing the state with an English summary, or escalating to a human when confidence is low.

Budgets and limits

• 64k tokens per request: covers state plus all questions.
• 32k tokens: covers state plus the single longest question.
• The model reads the state once, then evaluates every question in parallel, so packing many questions into one request costs almost nothing in latency — see Speculative fan-out.
• Accuracy shifts as the state grows; the official Jev 1.13 jaggedness section covers this.

Related

• Primitives — structuring questions with instructions and criteria
• Confidence — using the return value to control behaviour
• API reference — the request schema