Jev System One Models: Typed Decisions for Code
TypeSafe AI released Jev on Sep 15, a System One model that returns typed decisions with confidence. What RLCD changes, and how to wire thresholds in code.
On this page
Chat models write sentences, software needs answers it can branch on. On September 15, 2026 TypeSafe AI opened early access to Jev, a System One model that takes state plus typed questions and returns decisions with confidence. Here is what changes, how the three primitives work, and the threshold pattern I would ship first.
Takeaways
Jev trades strings for typed calls: state in, Choice plus Score plus Noul out, all evaluated in parallel. RLCD replaces preference tuning: every Choice and Score ships with confidence, so code can auto act or ask for review. Speed and price target automation: 70 to 500ms per call, $42 per billion input tokens, output too cheap to meter. Builder rule: decompose into atomic questions, combine in code, change weights without rewriting prompts.
What does Jev actually return?
Not text. You send one state string plus a map of questions, each with type, instructions, and criteria. Jev evaluates each question in isolation against the same state in one request, so adding questions barely changes latency and avoids context rot.
state: "Stripe integration failing for 3 days, losing sales, help ASAP"
questions: department = Choice(billing | technical | sales)
frustration = Score(0 calm | 1 civil | 2 angry)
is_urgent = Noul("conveys urgency")
answers: department technical 0.85, frustration 1.0, is_urgent 1.0The three primitives
Choice picks one option with probabilities and confidence. Score grades against a rubric with the same shape. Noul answers true or false as 0 to 1. Mix all three in a single call to /v1/systemone with model: jev-latest.
How do confidence thresholds work in code?
Ask atomic questions
Split "rate this pitch" into market size, feasibility, differentiation. One gut check per question keeps each judgment reliable and testable.
Set act and review bands
Above 0.85 auto act, 0.6 to 0.85 log plus continue, below 0.6 escalate. Thresholds live in code, not prompts, so policy changes ship as config.
Combine with your own weights
Multiply probabilities by your coefficients for routing or scoring. When priorities shift, edit one number instead of rewording a chain of thought.
Verify LLMs with Jev
Score LLM drafts, traces, and tool calls for jailbreaks, quality, and leakage. Small fast checks around big slow generators is the cheapest reliability win.
from typesafe_sdk import Choice, Noul, Score, TypeSafeClient
client = TypeSafeClient()
res = client.system_one(
state=ticket,
questions={
"department": Choice(instructions="Which team should handle this",
criteria={"billing": "Payment issues", "technical": "Bugs or integrations", "sales": "Pricing questions"}),
"frustration": Score(instructions="How frustrated the customer appears",
criteria=["Calm, just stating facts", "Frustrated but civil", "Very angry, strong language"]),
"is_urgent": Noul(instructions="The message conveys urgency"),
},
)
dept = res.answers["department"]
if dept.confidence > 0.85:
route(dept.choice)
else:
review_queue.add(ticket, dept.probabilities)Where does Jev fit, and where does it not?
Jev fits smart if statements, map reduce over data, real time UX under 500ms, and guardrails around LLM output. The Doom demo at 10 queries per second and Wikiracing over hundreds of links both show the same point: parallel decisions compound when hallucinations would break a chain. It does not fit chat, long reasoning, or code writing: no strings means no essays, no proofs, no refactors.
What should you verify before trusting the numbers?
Are 193.6x faster and 444.6x cheaper real?
That pair comes from TypeSafe workflow evals against the average of GPT-6 Astra and Fable 5.1 as reference, on four production style workflows. Treat it as the high end: simpler calls show smaller gaps, and evals ran from US West Coast laptops.
Is zero hallucination literally true?
For schema shape, yes by construction: outputs match the declared type or the call fails. For judgment quality, no: Jev can still pick the wrong choice with high confidence, which is why thresholds and review bands matter.
As of September 21, 2026: automation needed decisions with prices software can afford, not longer chat. Try one Choice plus one Noul on your own queue, set two thresholds, and keep the LLM for words. Next, read agents that do work for where those decisions now run for families and teams.