Jev (TypeSafe System One)

Introduction

• Jev is TypeSafe’s System One model — named after Kahneman’s Thinking, Fast and Slow: fast, structured decisions instead of open-ended generation.
• You provide a state (facts needed for the decision) and a set of questions; the API returns typed, calibrated answers ready for business logic — not prose to parse.
• Jev runs as a hosted API (closed weights). For a local open-weight alternative in the same niche, see GLiNER2.5-Decide and the QQP fine-tuning walkthrough.

Note

TypeSafe’s launch post: Introducing System One models and Jev.

Question types

Type	Role	Typical output
choice	Pick one of several options	Discrete label
score	Numeric rating on a rubric	Calibrated score
noul	Yes/no as a probability	Class and probability in \([0, 1]\) (threshold at 0.5)

Instructions for each question are natural-language criteria; the model scores against them without generating a chain-of-thought essay.

Architecture (conceptual)

System One / Jev-style decision flow (state + typed questions → calibrated answers)

• Unlike chat LLMs, Jev is built for decisions over a finite output space defined per request (choices, scores, or noul probabilities).
• Internals are not public; treat it as an API-first structured decision service rather than an inspectable encoder checkpoint.

Minimal example (duplicate detection)

from typesafe_sdk import Noul, TypeSafeClient

client = TypeSafeClient()

response = client.system_one(
    state={
        "question1": "How do I reset my password?",
        "question2": "What is the process to change my password?",
    },
    questions={
        "is_duplicate": Noul(
            instructions="Do question1 and question2 ask the same thing?",
        ),
    },
)

print(response.answers["is_duplicate"].noul)
# e.g. 0.71 → treat as duplicate if >= 0.5

When to use Jev

• Strong zero-shot structured decisions when you are fine with network latency and API billing (input tokens only on current pricing).
• No local GPU or no appetite to fine-tune an open encoder.
• Privacy allows sending state to a third-party API.

Compare latency and local deployment tradeoffs in the GLiNER2.5-Decide QQP blog.

References

[1] Jev — Product site | TypeSafe blog

[2] Related local model — GLiNER2.5-Decide