Decisions

Answer a list of typed questions about shared evidence and get one typed answer per question. This is OpenAI's Decisions endpoint (POST /v1/decisions, model gpt-6-luna), and it is also served by the decision models the proxy already reaches through System One — the two endpoints are one capability in two envelopes, and the proxy translates between them.

POST/v1/decisions

Request

bash

curl http://localhost:8080/v1/decisions \
  -H "Authorization: Bearer $LLM_PROXY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "input": "I was charged twice for my order.",
    "questions": [
      {
        "type": "predicate",
        "name": "urgent",
        "instructions": "Does this customer need a reply today?"
      },
      {
        "type": "choice",
        "name": "department",
        "instructions": "Which department should handle this complaint?",
        "choices": [
          {"value": "billing", "description": "Payments, invoices, and refunds."},
          {"value": "technical", "description": "Problems using the product."},
          {"value": "other", "description": "Requests outside these categories."}
        ]
      },
      {
        "type": "score",
        "name": "severity",
        "instructions": "How severe is this issue?",
        "levels": [
          {"label": "Cosmetic", "description": "Appearance only; no lost functionality."},
          {"label": "Workaround available", "description": "A task fails, but another way works."},
          {"label": "Fully blocked", "description": "A task fails with no workaround."}
        ]
      }
    ]
  }'

Response

json

{
  "model": "gpt-6-luna",
  "answers": [
    {"type": "predicate", "name": "urgent", "probability": 0.92},
    {
      "type": "choice",
      "name": "department",
      "choice": "billing",
      "probabilities": [
        {"value": "billing", "probability": 0.95},
        {"value": "technical", "probability": 0.02},
        {"value": "other", "probability": 0.03}
      ],
      "confidence": 0.93
    },
    {
      "type": "score",
      "name": "severity",
      "score": 1.1,
      "probabilities": [
        {"value": 0, "label": "Cosmetic", "probability": 0.1},
        {"value": 1, "label": "Workaround available", "probability": 0.7},
        {"value": 2, "label": "Fully blocked", "probability": 0.2}
      ],
      "confidence": 0.55
    }
  ],
  "usage": {
    "input_tokens": 296,
    "input_tokens_details": {"cache_write_tokens": 0, "cached_tokens": 0},
    "output_tokens": 0,
    "output_tokens_details": {"reasoning_tokens": 0},
    "total_tokens": 296
  }
}

Request fields

input (required) is the shared evidence every question reads: a plain string, or an array of user messages holding input_text and input_image parts. Images must be inline base64 data URLs (data:image/png;base64,…); hosted HTTP(S) URLs and file ids are not supported. At most 128 image parts are allowed across all messages in one request.

model (required) selects the decision model. Point the model at a provider whose type supports evaluation — openai for the Decisions endpoint itself, or typesafe, openrouter and ollama for the System One models the proxy bridges onto it.

questions (required) is the ordered list of questions; answers come back in the same order. Every question has a type and instructions, and may carry a name — a string when given (explicit null is rejected) — that the answer echoes. Names must be unique, and a name starting with __unnamed_question_ is rejected: the proxy uses that prefix to key unnamed questions when translating for a System One upstream, so a client name under it would collide with one. An unnamed question answers with "name": null, the shape the endpoint itself emits.

`type`	Shape	Extra fields
`predicate`	Yes/no question; returns the probability the condition is true	—
`choice`	Picks one option; returns the option and the full distribution	`choices`: non-empty list of `{value, description?}` (at most 255). `value` is a string or a boolean — `true` and `"true"` are distinct options
`score`	Rates along a rubric; returns a probability-weighted value	`levels`: 2–26 ordered `{label, description?}`, lowest first

safety_identifier is accepted and forwarded to upstreams that document it (openai); the System One upstreams have no equivalent field and drop it.

Unknown top-level fields are rejected with a 422.

One endpoint, two envelopes

System One and Decisions answer the same three primitives. The proxy keeps both endpoints available on every evaluation model and translates at the provider:

	Decisions	System One
Evidence	`input` (string or messages)	`state` (string, object or array)
Questions	ordered array, each with a `name`	map of id to question
Yes/no	`predicate`	`noul`
Pick one	`choices[{value, description}]`	`criteria{value: description}`
Rate	`levels[{label, description}]`	`criteria[level, …]`
Answers	ordered array, echoing `name`	map keyed by question id

So a request to /v1/decisions naming a Jev model is converted to a System One request, sent to /v1/systemone, and its answers converted back — and a request to /v1/systemone naming a gpt-6-luna model takes the reverse path. The conversion is as lossless as the two envelopes allow; see Cross-endpoint losses for what each direction cannot carry.

Limits

The proxy rejects a malformed request up front with a 422 naming the offending question. Where the upstreams disagree on a limit, the proxy accepts the union and lets the upstream that does not support the request answer for itself:

score takes 2–26 ordered levels — Ollama's ceiling. TypeSafe and OpenRouter stop at 10, so a rubric above that is served by Ollama and rejected upstream by the other two.
choice takes at most 255 options.
Question names must be unique when given, and choice option values must stay distinct after a boolean is written as its word: the bridge keys a System One criteria map by string, where true and "true" would collapse into one option.

Response

answers carries one typed answer per question, in question order. Every answer has a type — predicate, choice, score, or refusal (the host declined one question without disclosing a score).

predicate → probability, 0 to 1
choice → choice (a supplied value, string or boolean), probabilities for every option, and confidence
score → score (the probability-weighted average of the level indices, so it can fall between levels), probabilities with each level's value and label, and confidence

model identifies the model that performed the evaluation and usage reports input and output tokens. There is no streaming variant: a decision is a single synchronous evaluation. usage is echoed verbatim from the upstream when it reported one.

An unnamed question has no name in its answer; correlate it by position. A bridged answer omits confidence when the System One upstream reported none.

Provider support

The model named in the request must be marked as a decision model (supports_systemone) in the admin model configuration; a chat model is rejected with a 400 invalid_request_error before any provider is called. The same flag gates /v1/systemone and the routing judge.

Provider type	Upstream endpoint	Notes
`openai`	`POST /v1/decisions`	Native. The only provider that accepts inline images for a decision.
`typesafe`	`POST /v1/systemone`	Bridged. Images are dropped (no image channel upstream).
`openrouter`	`POST /v1/systemone`	Bridged. OpenRouter serves the same envelope on a second, still-alpha route (`/api/alpha/decisions`); this stays on the stable one. Images are dropped.
`ollama`	`POST /v1/systemone`	Bridged. Inline images become the local `images` field (raw base64, as Ollama documents).

Typesafe and Ollama serve one route for both endpoints, so an endpoint_base_urls.systemone override moves both. OpenRouter names a second route for the same envelope, so its /v1/decisions resolves under its own endpoint_base_urls.decisions key, which is how a deployment opts into the alpha route without moving /v1/systemone:

YAML

yaml

endpoint_base_urls:
  decisions: "https://openrouter.ai/api/alpha/decisions"

Cross-endpoint losses

Both directions are lossless except where one envelope has no field for a fact:

Decisions → System One

A predicate has no rubric, so the System One noul criteria map is left empty rather than invented.
score level descriptions move into the question text (as a numbered rubric) because System One's criteria array holds the level strings the answer legend echoes. The labels the client sent come back unchanged.
input messages are flattened to one text block — every message is a user message, so the evidence keeps its order — and their inline images move to the local images field, which only ollama documents. safety_identifier has no System One counterpart and is dropped.

System One → Decisions

A noul rubric (criteria.true/criteria.false) is folded into the instructions, since a predicate has no rubric field.
A structured state (object or array) becomes JSON text, and images become inline input_image parts whose media type is read from the image's leading bytes. The OpenRouter-only fields (provider, session_id, trace, user) and Ollama's keep_alive have no Decisions counterpart and are dropped.

System One evaluation — the sibling envelope
Endpoint Index
Errors & Rate Limits

Request

bash

curl http://localhost:8080/v1/decisions \
  -H "Authorization: Bearer $LLM_PROXY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "input": "I was charged twice for my order.",
    "questions": [
      {
        "type": "predicate",
        "name": "urgent",
        "instructions": "Does this customer need a reply today?"
      },
      {
        "type": "choice",
        "name": "department",
        "instructions": "Which department should handle this complaint?",
        "choices": [
          {"value": "billing", "description": "Payments, invoices, and refunds."},
          {"value": "technical", "description": "Problems using the product."},
          {"value": "other", "description": "Requests outside these categories."}
        ]
      },
      {
        "type": "score",
        "name": "severity",
        "instructions": "How severe is this issue?",
        "levels": [
          {"label": "Cosmetic", "description": "Appearance only; no lost functionality."},
          {"label": "Workaround available", "description": "A task fails, but another way works."},
          {"label": "Fully blocked", "description": "A task fails with no workaround."}
        ]
      }
    ]
  }'

Response

json

{
  "model": "gpt-6-luna",
  "answers": [
    {"type": "predicate", "name": "urgent", "probability": 0.92},
    {
      "type": "choice",
      "name": "department",
      "choice": "billing",
      "probabilities": [
        {"value": "billing", "probability": 0.95},
        {"value": "technical", "probability": 0.02},
        {"value": "other", "probability": 0.03}
      ],
      "confidence": 0.93
    },
    {
      "type": "score",
      "name": "severity",
      "score": 1.1,
      "probabilities": [
        {"value": 0, "label": "Cosmetic", "probability": 0.1},
        {"value": 1, "label": "Workaround available", "probability": 0.7},
        {"value": 2, "label": "Fully blocked", "probability": 0.2}
      ],
      "confidence": 0.55
    }
  ],
  "usage": {
    "input_tokens": 296,
    "input_tokens_details": {"cache_write_tokens": 0, "cached_tokens": 0},
    "output_tokens": 0,
    "output_tokens_details": {"reasoning_tokens": 0},
    "total_tokens": 296
  }
}

YAML

yaml

endpoint_base_urls:
  decisions: "https://openrouter.ai/api/alpha/decisions"

Decisions ​

Request fields ​

One endpoint, two envelopes ​

Limits ​

Response ​

Provider support ​

Cross-endpoint losses ​

Related ​

Decisions

Request fields

One endpoint, two envelopes

Limits

Response

Provider support

Cross-endpoint losses

Related