Migrating a System One client

decisio serves POST /v1/systemone and GET /v1/models in TypeSafe's System One wire format, so an existing client needs only a new base URL and a model name. This page shows how to point a client at a decisio server and lists the differences that matter once it is there. The SDK steps below were run with TypeSafe's Python SDK 0.7.2 against a decisio server on 2026-10-05. The SDK is TypeSafe's, published under the MIT licence (PyPI, repository); decisio does not include it, and a client installs it on its own. decisio is an independent project, not affiliated with or endorsed by TypeSafe; it implements TypeSafe's published System One wire format.

Point the client at the server

With TypeSafe's Python SDK, pass the server's address and its served model name:

from typesafe_sdk import TypeSafeClient

client = TypeSafeClient(
    base_url="http://127.0.0.1:8000",
    api_key="unused",                          # the SDK requires one; decisio ignores it
    model="decisio-qwen3.6-35b-a3b-letters",   # the name GET /v1/models lists
)
answer = client.system_one(
    state="Since this morning none of our staff can log in to the dashboard.",
    questions={"category": {"type": "choice", "instructions": "Which team should handle this ticket?",
                            "criteria": {"billing": None, "access": None, "bug": None}}},
)

Run against the CPU stand-in at v0.8.2 on 2026-10-07: with TypeSafe's Python SDK 0.7.2, the version the page names.

Or leave the code as it is and set the SDK's environment variables:

export TYPESAFE_BASE_URL=http://127.0.0.1:8000
export TYPESAFE_API_KEY=unused
export TYPESAFE_DEFAULT_MODEL=decisio-qwen3.6-35b-a3b-letters

Run against the CPU stand-in at v0.8.2 on 2026-10-07: sets the variables the SDK reads.

Any other client sends the same JSON body to http://<host>:8000/v1/systemone. client.models.list(), or curl http://127.0.0.1:8000/v1/models, returns the served name: decisio-qwen3.6-35b-a3b-letters, decisio-gemma-4-12b-it-letters or decisio-gemma-4-31b-it-letters, unless the server was started with --served-name.

The differences that matter

The model field does not choose the model. A decisio server serves one base, chosen when it starts (--base), whatever the request names. The response echoes the name the request sent, so a client left on the SDK's default model name gets that name back unchanged. Set the client's model to the served name, and read GET /health (its base and profile) when you need to know what answered.

There is no authentication. The Authorization header is accepted and ignored, and so is any API key. Keep the server on 127.0.0.1, or behind a reverse proxy that authenticates (SECURITY.md).

Answers are read, not generated. Each question's distribution comes from the option letters' logits at one position, in one forward pass, tempered by the base's fitted temperatures. confidence follows System One's rules, usage.output_tokens is one per question, and usage.input_tokens counts tokens in the base's own tokenizer, so token counts differ from another provider's.

Unknown request fields are ignored, not rejected. A field decisio does not know is accepted and has no effect, so check that a field your client relies on is in docs/api.md. The extensions decisio reads are images and orders.

Limits. Up to 255 options per question (a request with more gets 400), a prompt of up to 32,768 tokens, and up to 2 images per request, which need a server started with --image-model.

Errors. A request that fails validation gets 422 and one the engine cannot answer gets 400, each with a detail naming the problem. TypeSafe's SDK raises its own error types for them (TypeSafeUnprocessableEntityError, TypeSafeBadRequestError). decisio sets no rate limit, so the SDK's retries on 429 never fire; it still retries 408 and 5xx.

Timeouts. The SDK's default timeout is 10 s. A long state seen for the first time takes longer on the larger bases, and the first request after a start warms the engine; raise timeout for those.

Repeatability is stated per base. On the Qwen base a question returns the same probabilities every time, alone or among other questions; the Gemma 4 12B base returns the same choice with probabilities that can move slightly; EVAL_CARD.md sections 6.5 and 7.5 have what was measured on each Gemma base (and the README's "Choosing a base", docs/cli.md "Determinism"). Among exactly tied options, the answer is the key that sorts first, so it never depends on the order the keys arrive in.

What decisio adds.

• Per-task calibration and intent heads, learned from your own labelled examples (POST /v1/tasks, docs/tasks.md).
• Per-task abstention thresholds (POST /v1/abstention/tasks).
• Response headers naming the server's time, the route and the tasks applied (x-decisio-server-ms, x-decisio-route, x-decisio-tasks, x-decisio-stages).
• GET /health, naming the base, its checkpoint and revision, the temperatures and the prompt.

docs/api.md has every field, and docs/running.md every way to start a server.