Back to changelog

System One: Typed Decisions

The new /v1/systemone endpoint returns typed decisions with calibrated probabilities instead of generated text. TypeSafe's Jev 1.13 is the first decision model, billed on input tokens only.

A glowing decision switch on a circuit-board chip surrounded by probability dials and branching paths, representing typed decisions on LLM Gateway

Plenty of model calls exist only to make a decision: route this ticket, score this passage, check whether this citation supports the claim. A chat model can do it, but you then parse prose, hope the JSON holds, and get no number to threshold on. System One is a new endpoint for models that skip the text entirely — you name the questions, and every answer comes back as a value your code can branch on, with a probability attached.

One endpoint, three question types

POST https://api.llmgateway.io/v1/systemone takes a state and a map of questions keyed by ids you choose. Answers come back under the same ids.

Type Ask Answer
noul A yes/no question noul: probability the answer is yes, 0 to 1
choice One option from a set you define choice, probabilities per option, confidence
score A rating across ordered levels score (can land between levels), legend, confidence 
1curl -X POST "https://api.llmgateway.io/v1/systemone" \2  -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "model": "jev-1.13.0",6    "state": "Our production integration has been returning 500s for 3 days.",7    "questions": {8      "department": {9        "type": "choice",10        "instructions": "Which team should handle this?",11        "criteria": {12          "billing": "Payments, invoicing, refunds",13          "technical": "Bugs, outages, integrations",14          "sales": "Pricing, upgrades, new accounts"15        }16      },17      "is_urgent": {18        "type": "noul",19        "instructions": "Does this convey urgency?"20      },21      "impact": {22        "type": "score",23        "instructions": "Rate the operational impact.",24        "criteria": ["None", "Limited", "Critical"]25      }26    }27  }'

state and every instructions field accept structured data as well as plain text, so a question can reference fields of the record you passed in. Choice questions take up to 255 options; score questions take two to ten ordered levels.

The policy stays in your code: threshold noul where you want the cutoff, and use confidence as a second axis to send uncertain cases to a human instead of acting on a coin flip.

Jev 1.13, billed on input only

The first decision model is typesafe/jev-1.13.0 from TypeSafe AI, at $0.042 per million input tokens with output free — the response is a set of values, not generated text. Moving aliases such as jev-latest resolve to the pinned version, so a request is always billed and logged against a version you can read prices for.

Because the state is read once and every question is evaluated against it, a batch of questions costs far less than one request each — including speculative questions you only read when another answer makes them relevant. Question text counts as input on every call, so a large rubric has a real per-request cost.

Decision models only work on /v1/systemone. Asking for one on /v1/chat/completions returns a 400 pointing at the right endpoint; they cannot generate text or call tools, and they are not in the Playground. Input is text only. The endpoint runs the same machinery as every other gateway route: IAM rules, compliance policies, credit gating, provider key rotation, request logging with a per-model cost breakdown, and a per-organization rate limit of 600 requests per minute.

System One docs → | Browse decision models →