jevper¶

OpenAI's Decisions API answers from any OpenAI-compatible model. client.decisions.create(input=…, questions=[…]) takes the documented Decisions request and returns its documented response — model, answers in question order, and usage — evaluated by whatever model you point it at, over Chat Completions, Responses, or the Anthropic-compatible Messages API. That is what to reach for where POST /v1/decisions is not served yet: an endpoint without the route, a model the beta does not take (it serves gpt-6-luna alone), a local server, or a gateway in front of any of them. The models you already run answer the same questions here, in the same schema. See Decisions API compatibility.

Underneath, jevper is a typed-question library: a state and noul, choice, or score questions in, typed answers out — choice and score answers carrying probabilities and confidence, a noul answer its noul. The Jev (System One) wire format is that library's own interface, and the whole of it is documented here; client.system_one(...) is the call the Decisions wrapper is built on.

jevper is an independent implementation of OpenAI's documented Decisions API schema and of the documented System One wire format. It is not affiliated with, endorsed by, or supported by OpenAI or TypeSafe AI. The package does not call the hosted TypeSafe API and does not import an OpenAI or Anthropic SDK at runtime.

Install¶

Python 3.10 or newer is required. Install the package from PyPI:

pip install jevper

This site documents jevper 0.7.12. python -c "import jevper; print(jevper.__version__)" says which release your environment has; a different number means the site is ahead of, or behind, what you installed.

The only runtime dependency is pydantic>=2.7. You provide a client object; jevper does not read credentials or provider configuration from the environment.

First call¶

from openai import OpenAI
from jevper import SystemOneClient

client = SystemOneClient(OpenAI(), model="gpt-5.6-terra")

decision = client.decisions.create(
    input="I was charged twice for the same subscription this month.",
    questions=[
        {
            "type": "choice",
            "name": "intent",
            "instructions": "Pick the intent of the message.",
            "choices": [
                {"value": "billing", "description": "money, invoices, refunds, charges"},
                {"value": "technical", "description": "errors, crashes, login or performance problems"},
                {"value": "sales", "description": "pricing, plans, purchasing, upgrades"},
            ],
        }
    ],
)

answer = decision.answers[0]
print(answer.name, answer.choice, [entry.value for entry in answer.probabilities], answer.confidence)

method="auto" is the default: jevper uses logprobs when the provider returns a usable distribution and falls back to structured output when it does not. Whatever it chose, and what each provider call cost, is in decision.raw.debug and decision.raw.usage — the jevper response the wrapper converted, kept beside it and left out of every dump.

The Jev interface underneath¶

The same call without the Decisions schema is system_one, which returns the answers keyed by your own question ids:

from jevper import Choice, SystemOneClient

client = SystemOneClient(OpenAI(), model="gpt-5.6-terra")

response = client.system_one(
    state="I was charged twice for the same subscription this month.",
    questions={
        "intent": Choice(
            instructions="Pick the intent of the message.",
            criteria={
                "billing": "money, invoices, refunds, charges",
                "technical": "errors, crashes, login or performance problems",
                "sales": "pricing, plans, purchasing, upgrades",
            },
        )
    },
)

answer = response.answers["intent"]
print(answer.choice, answer.probabilities, answer.confidence)

Either response converts to the other shape: response.to_decision() returns the Decisions schema for a system_one response, with the names, values and labels the response itself remembers. See API reference.

Choose a page¶

How a call is selected¶

flowchart LR
    A["State and typed questions"] --> B["Validate and render prompt"]
    B --> C{"method and surface"}
    C -->|logprobs| D["Provider logprobs"]
    C -->|grammar| E["Constrained label"]
    C -->|structured or discrete| F["JSON answer"]
    D --> G["Read distribution"]
    E --> G
    F --> G
    G --> H["Typed answer, usage, and debug data"]

Start with Getting started for the smallest working path, then use Methods and Local servers to tune the provider surface.