Decision model: Eve

Eve (eve) is the Wally Cloud decision model. It does not chat and generates no text: it answers yes/no, choice and score questions about an input with a probability for each answer. Every question is scored on its own, in prefill, and each answer’s probabilities sum to 1. From the terminal, wally decide (or wally decisions) uses Eve by default, so you don’t need to pass a model.

Endpoints

POST https://inference.runanywhere.ai/v1/decisions
POST https://inference.runanywhere.ai/v1/systemone
Authorization: Bearer sk-runa-…

POST /v1/decisions is the Wally request shape, documented on this page.
POST /v1/systemone takes the same questions in the System One shape that OpenRouter’s Decisions API and TypeSafe use. It is served by the same handler, with the same billing and limits. See System One below.
Eve is not listed by GET /v1/models. Call it by its id.
Chat and text completions on eve are refused with 400 unsupported_endpoint.

Authentication is the same Cloud key as chat, created in console.runanywhere.ai. The examples read it from RUNA_CLOUD_KEY.

Request

POST /v1/decisions takes a JSON body with these fields and no others. A body with any other field is refused with 400.

Field	Type	Required	Allowed values and limits
`model`	string	yes	`eve`
`input`	string	yes	The text every question is asked about. Not blank. It is rendered into each question’s prompt, so it is billed once per question.
`questions`	array	yes	1 to 128 questions, each with an `id` unique in the request. See below.
`temperature`	number	no	Greater than 0, at most 100. Default 1. Divides the label log-probabilities before the softmax, so it changes `probabilities` only.
`prompt_format_version`	integer or null	no	Pins the server’s prompt wording. Eve serves version `3`. Any other version is refused with `400` before anything is billed. Null or absent means the served version.

Questions

Every question has these fields:

Field	Type	Required	Limits
`id`	string	yes	1 to 128 characters, not blank. The answer comes back under it.
`type`	string	yes	`yes_no`, `choice` or `score`
`question`	string	yes	Not blank, at most 65,536 characters

Then, by type:

Type	Extra fields	Labels in the answer
`yes_no`	`yes`, `no` (optional): text for the yes and the no answer	`yes`, `no`
`choice`	`options` (required): 2 to 26 objects, each `{ "name", "description"? }`	the option names
`score`	`levels` (required): 2 to 10 strings, labelled 0 to 9 in the order given	the level indices, `"0"` to `"9"`

An option name is a single line of at most 256 characters, and names must be distinct within a question, ignoring case and surrounding space.

Limits

The whole body is at most 1 MiB.
Each question’s prompt (the input, that question’s wording and the server’s template) must fit in 8,185 tokens. A longer one is refused with 400, and nothing is billed.
Per key: at most 32 requests in flight and 600 requests a minute. Over either, the gateway answers 429 with Retry-After.

One request per question type

A yes/no question:

{
  "model": "eve",
  "input": "After the 2.3 deploy, every order confirmation email shows the wrong currency for EU customers.",
  "questions": [{ "id": "is_bug", "type": "yes_no", "question": "Is this a bug report?" }]
}

A choice question:

{
  "model": "eve",
  "input": "After the 2.3 deploy, every order confirmation email shows the wrong currency for EU customers.",
  "questions": [
    {
      "id": "team",
      "type": "choice",
      "question": "Which team owns this?",
      "options": [{ "name": "frontend" }, { "name": "payments" }, { "name": "email" }]
    }
  ]
}

A score question:

{
  "model": "eve",
  "input": "After the 2.3 deploy, every order confirmation email shows the wrong currency for EU customers.",
  "questions": [
    {
      "id": "severity",
      "type": "score",
      "question": "How severe is it?",
      "levels": ["cosmetic", "minor", "major", "outage"]
    }
  ]
}

Response

A 200 has one answer per question, keyed by the question’s id.

Field	Type	Meaning
`object`	string	Always `decisions`
`model`	string	The model id you sent
`prompt_format_version`	integer	The prompt wording the answers were computed under (`3` for Eve)
`answers.<id>.type`	string	`yes_no`, `choice` or `score`
`answers.<id>.probabilities`	object	One probability per label, summing to 1
`answers.<id>.label_mass`	number	The share of the model’s next-token distribution that landed on the labels. Always 1 for Eve.
`answers.<id>.choice`	string	Choice questions only: the most probable option name
`answers.<id>.score`	number	Score questions only: the probability-weighted level, from 0 to 9
`usage.prompt_tokens`	integer	Tokens prefilled: the input counted once per question, plus each question’s wording
`usage.completion_tokens`	integer	Always 0
`usage.reasoning_tokens`	integer	Always 0
`usage.total_tokens`	integer	Equal to `prompt_tokens`

Example

The three questions above, in one request:

curl https://inference.runanywhere.ai/v1/decisions \
  -H "Authorization: Bearer $RUNA_CLOUD_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "eve",
    "input": "After the 2.3 deploy, every order confirmation email shows the wrong currency for EU customers.",
    "questions": [
      { "id": "is_bug", "type": "yes_no", "question": "Is this a bug report?" },
      {
        "id": "team",
        "type": "choice",
        "question": "Which team owns this?",
        "options": [{ "name": "frontend" }, { "name": "payments" }, { "name": "email" }]
      },
      {
        "id": "severity",
        "type": "score",
        "question": "How severe is it?",
        "levels": ["cosmetic", "minor", "major", "outage"]
      }
    ]
  }'

The response, from a production call on 2026-10-04:

{
  "object": "decisions",
  "model": "eve",
  "prompt_format_version": 3,
  "answers": {
    "is_bug": {
      "type": "yes_no",
      "probabilities": {
        "yes": 0.9902915212263165,
        "no": 0.009708478773683474
      },
      "label_mass": 1.0
    },
    "team": {
      "type": "choice",
      "probabilities": {
        "frontend": 0.02328114907703061,
        "payments": 0.41266824142534275,
        "email": 0.5640506094976266
      },
      "label_mass": 1.0,
      "choice": "email"
    },
    "severity": {
      "type": "score",
      "probabilities": {
        "0": 0.06803281960131544,
        "1": 0.16838284565086212,
        "2": 0.7089184284464499,
        "3": 0.05466590630137255
      },
      "label_mass": 1.0,
      "score": 1.7502174214478796
    }
  },
  "usage": {
    "prompt_tokens": 317,
    "total_tokens": 317,
    "completion_tokens": 0,
    "reasoning_tokens": 0
  }
}

Errors

Errors use the same envelope as chat: { "error": { "message", "type", "code", "param" } }. A refused request is not billed.

Status	`code`	When
`400`	`bad_request`	The body breaks the schema (an unknown field, a missing field, too many questions or options), a question is over the 8,185-token window, or `prompt_format_version` is not `3`
`400`	`unsupported_endpoint`	A chat or text completion on `eve`, or a decision on a model that is not a decision model
`401`	`invalid_api_key` and related	The key is missing, unknown, expired or revoked
`403`	`model_not_entitled`	The key cannot call this model
`429`	(the limiter’s)	The key is over 32 in flight or 600 requests a minute. Retry after `Retry-After` seconds.
`503`	`draining`	The model is briefly unavailable. Retry after `Retry-After` seconds.
`504`	`timeout`	No answer in time

Billing

A decision bills its input once per question at the input rate. Nothing is generated, so there are no output tokens to bill. At pricing version 2026-10-03.4, Eve’s rate is $0.042 per million input tokens. The console shows current rates for every model.

System One

POST /v1/systemone answers the same questions in the System One shape. The model, limits and billing are the same; only the request and response shapes differ:

`/v1/decisions`	`/v1/systemone`
`input`: a string	`state`: a string, or a JSON array or object rendered as compact JSON
`questions`: an array, each with `id`	`questions`: an object keyed by your own names (1 to 128)
`yes_no`, `choice`, `score`	`noul` (yes/no), `choice`, `score`
`question`	`instructions`
`yes`/`no`, `options`, `levels`	`criteria`: `{ "true", "false" }` for noul, `{ "<label>": guidance }` for choice, an array of level descriptions for score
`temperature`, `prompt_format_version`	Not accepted (refused unless null). Other unknown top-level fields are ignored.
Response `answers.<id>.probabilities`	Response `answers.<name>`: `noul` (the probability of yes); `choice`, `confidence` and `probabilities`; `score`, `confidence`, `probabilities` and `legend`
`usage.prompt_tokens`	`usage.input_tokens` and `usage.output_tokens` (always 0), plus a top-level `id` equal to the `x-request-id` header

The full schemas for both routes are in the Wally Decisions API contract, version 1.1.0.

Get started

API reference

Reference

Endpoints

Request

Questions

Limits

One request per question type

Response

Example

Errors

Billing

System One

​Endpoints

​Request

​Questions

​Limits

​One request per question type

​Response

​Example

​Errors

​Billing

​System One

Endpoints

Request

Questions

Limits

One request per question type

Response

Example

Errors

Billing

System One