Reference · TypeSafe Jev

Glossary & primitives cheat-sheet

The compressed source of truth. TypeSafe's exact nomenclature and the request/response shapes every lesson leans on. Print it; keep it beside the editor.

ReferencePrints on one–two pages

Core nomenclature

Term	Meaning
System One model	A model built to make fast, structured decisions software consumes directly — as opposed to a text-generating "System 2" chat model. TypeSafe's category.
Jev	TypeSafe's first public System One model, optimised for automation. The thing you actually call. Version strings look like `jev-1.13.0`; request with `jev-latest`.
State	Everything you'd put in front of a panel of experts before asking them to judge — the text/JSON context. Sent once, shared by every question in the call.
Question	One typed ask against the state. Has a `type` (primitive), `instructions`, and (for Choice/Score) `criteria`. Questions are a map, keyed by names you choose.
Primitive	The type of a question: `choice`, `score`, or `noul`. Modular, composable, structured, fast.
Probabilities	The model's distribution over the possible answers — how the confidence is spread across options/levels.
Confidence	A single number for how sure the model is overall (Choice & Score). Your code gates on it. Noul has no separate confidence — the 0–1 value is the signal.

The three primitives

Primitive	Asks	Answer fields
Choice	Pick one from a fixed set of named options.	`choice`, `probabilities` (per option), `confidence`
Score	Rate the state on an ordered rubric (levels you define).	`score`, `probabilities` (per level), `confidence`
Noul	Judge whether a statement about the state is true.	a single `probability`, 0–1

Request shape

POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY
Content-Type: application/json

{
  "model": "jev-latest",
  "state": "<text or JSON context to judge>",
  "questions": {
    "<your_key>": {
      "type": "choice | score | noul",
      "instructions": "<what to decide>",
      "criteria": { "<option_or_level>": "<description>", ... }
    }
  }
}

Response shape

{
  "model": "jev-1.13.0",
  "answers": {
    "<your_key>": {
      "type": "choice",
      "choice": "technical",
      "probabilities": { "billing": 0.08, "technical": 0.85, "sales": 0.07 },
      "confidence": 0.82
    }
  },
  "usage": { "input_tokens": 312, "output_tokens": 48 }
}

Golden rules 1. Decompose — many small atomic questions, combined in your code, beat one big one.
2. Gate on confidence — auto-act when high, escalate to a human when low.
3. One call, many questions — they run in parallel on shared state; latency barely moves.
4. Jev decides, it never writes — need words for a human? That's an LLM's job.

Sources

docs.typesafe.ai — official documentation (home, primitives)
Introduction — System One vs text, the mismatch
LiteLLM — TypeSafe pass-through — verified request/response wire format
OpenRouter — TypeSafe · Cloudflare AI — Jev

Back to the Jev course · Lesson 01