Decision Models

Probabilities, not text

A decision model reads some state and answers your typed questions with calibrated probabilities. Use it for routing, classification and yes/no gates inside an app, where a fast, structured answer matters more than prose.

Two decision models are served through the gateway, both on the /v1/systemone endpoint in the SystemOne format:

Model	Maker	Context	Input	Output
sage-2	Levanto Labs	32K	$0.07 / 1M	$2.00 / 1M
jev-1.13	TypeSafe	32K	$0.042 / 1M	free

They don't serve /v1/chat/completions or /v1/messages: sending one there returns a 400 that points to /v1/systemone. For the same reason, /v1/models/health shows them as unknown when they've had no calls in the last 15 minutes: idle models are checked with a chat probe, which a decision model can't answer, so that status doesn't mean an outage.

Asking a question

POST /v1/systemone

Authentication is the same as every other gateway endpoint: a Bankr API key (bk_...) via X-API-Key or Authorization: Bearer.

curl -X POST https://llm.bankr.bot/v1/systemone \
  -H "Content-Type: application/json" \
  -H "X-API-Key: bk_YOUR_API_KEY" \
  -d '{
    "model": "sage-2",
    "state": "User: please swap 1 ETH to USDC on Base",
    "questions": {
      "is_trade": {
        "type": "noul",
        "instructions": "Is the user asking to make a trade?"
      },
      "chain": {
        "type": "choice",
        "instructions": "Which chain does the user want to trade on?",
        "criteria": { "base": "Base", "ethereum": "Ethereum mainnet", "solana": "Solana" }
      }
    }
  }'

Request

Field	Description
model	Required. sage-2 or jev-1.13.
state	Required. What the questions are about: a string, or a JSON object.
questions	Required. An object of at least one question, keyed by an ID you choose.

Apart from the privacy fields and "stream": false, any other field is rejected with 400 unsupported_parameter. Streaming is not supported.

Every question has a type and instructions. The three types:

Type	Asks	criteria
noul	A yes/no question	None
choice	Pick one option	Required. An object of option ID to description
score	Where it sits on a scale	Required. A list of scale labels, from the lowest to the highest

Response

{
  "id": "dec-k5lcsV1k57JzQ2qjXrDUL",
  "model": "sage-2",
  "answers": {
    "is_trade": { "type": "noul", "noul": 0.991 },
    "chain": {
      "type": "choice",
      "choice": "base",
      "confidence": 0.9997,
      "probabilities": { "base": 0.9997, "ethereum": 0.0002, "solana": 0.00005 }
    }
  },
  "usage": { "input_tokens": 127, "output_tokens": 0 }
}

Each answer carries its question's type:

• noul: noul is the probability that the answer is yes.
• choice: choice is the most likely option, with its confidence and the probabilities of every option.
• score: score runs from 0 (the first label) to 1 (the last), with probabilities per label (by position) and a legend mapping each position to its label.

Privacy

Decision models are served without TEE attestation, so a private request ("privacy": "private", "private": true, sage-2:private or the /private path) returns 422 confidential_unavailable.

Zero data retention depends on the model:

• jev-1.13 takes a ZDR request ("privacy": "zdr", "zdr": true, jev-1.13:zdr or the /zdr path), which only reaches a provider that retains nothing. Accounts with zero data retention turned on can call it too.
• sage-2 has no zero-retention route, so a ZDR request returns 422 zdr_unavailable, and so does every Sage call from an account with zero data retention turned on.

See Privacy Tiers for how the tiers work on chat models.

Error responses

Status	Code	Meaning
400	missing_required_fields	model or state is missing.
400	invalid_value	questions is missing, empty or not an object.
400	unsupported_parameter	The body has a field other than the ones above.
400	stream_not_supported	stream was set to anything but false.
400	unsupported_model	The model is unknown or not a decision model.
400	invalid_privacy	A privacy field has a value it doesn't take.
400	privacy_conflict	The request asks for a tier the path doesn't.
422	confidential_unavailable / zdr_unavailable	Private tier, or ZDR with no zero-retention hosts.

A malformed question (for example a choice with no criteria) is rejected by the model's host with a 4xx, whose message usually names the field.

Pricing

Each call bills its input_tokens and output_tokens at the model's rates from your LLM credit balance, and shows on your usage dashboard. A short request costs a fraction of a cent: the 127-token call above costs about $0.00001 on Sage. See credit management to top up.

Next steps

• API Reference: endpoint summary
• Supported Models: the chat model catalog
• Overview: routing, payments and credit management