API· Last updated 2026-09-21

Jev API decision endpoint

The only first-party decision endpoint TypeSafe documents is:

POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer <API_KEY>
Content-Type: application/json

Jev (TypeSafe System One) evaluates one state against a map of typed questions and returns answers under the same ids. This unofficial page is a developer field table. Canonical prose: docs.typesafe.ai/api. We do not sell keys. Companion list: GET https://api.typesafe.ai/v1/models.

What you POST

Field	Required	Type	Notes from the official API reference
`state`	yes	string, object, or array of text	Material to judge. No image/audio/video.
`model`	yes	string	Examples use `jev-latest`. Pin `jev-1.13.0` when thresholds are fitted.
`questions`	yes	object	Map of your ids → question objects. Keys are not sent to the underlying model.

Each question: type (noul | choice | score) + instructions (string, object, or array) + type-specific criteria. Choice criteria: map, max 255 options. Score criteria: ordered array, 2–10 levels. Noul criteria: optional {true,false} glosses.

What you GET

Field	Role
`model`	Versioned id that answered (quick-start examples: `jev-1.13.0`) even if you sent an alias
`answers` map	Same keys you sent
`usage.input_tokens` / `usage.output_tokens`	Billing is input-only on the public models page (vendor claim: $0.042 / MTok; output free)

Answer shapes: see OpenAPI-like shape. HTTP errors: 401 / 422 / 429 / 529 — error envelope.

Gotcha: `jev-1.13` vs `jev-1.13.0`

String	Where it belongs
`jev-1.13.0`	First-party versioned id. Models page column for “Jev 1.13”. Pin this.
`jev-latest` / `jev-preview`	First-party aliases. Both pointed at `jev-1.13.0` on the models page we cited; `jev-preview` moves first when a preview exists.
`jev-1.13` (no patch)	Appears in some cookbook snippets and in OpenRouter as `typesafe/jev-1.13`. Do not assume first-party HTTP accepts the unpatched name.
`typesafe-ai/jev`, `typesafe/jev`	Gateway / Workers AI ids — wrong host if you POST to `api.typesafe.ai`.

GET /v1/models currently lists aliases. Versioned ids are still valid on POST whether or not they appear in that list (TypeSafe models page).

Dual-stack sketch

from typesafe_sdk import Noul, TypeSafeClient

client = TypeSafeClient()  # TYPESAFE_API_KEY, default jev-latest
r = client.system_one(
    state="Payouts failing for 3 days.",
    questions={"urgent": Noul(instructions="Does this convey urgency?")},
)
print(r.model, r.answers["urgent"].noul)

import { TypeSafeClient, noul } from "@typesafe-ai/sdk";

const client = new TypeSafeClient();
const r = await client.systemOne({
  state: "Payouts failing for 3 days.",
  questions: { urgent: noul({ instructions: "Does this convey urgency?" }) },
});

Method names follow current SDK docs — if they drift, believe docs.typesafe.ai/sdk.

FAQ

Is there a batch HTTP route? Not in the public API reference we used. Extra questions on one state are the documented parallelism. Treat “batch endpoint” inventory pages as application patterns unless TypeSafe adds a URL.

Chat completions? Wrong contract. Structured decisions are not a generated JSON document.

Hub: API. Siblings: request schema, response shape. Official: docs.typesafe.ai/api.

Sources

Public TypeSafe or adjacent documentation only. No private claims.