SYSTEM ONE MODEL · JEV

Not chat. Judgment.

A probabilistic model software can use directly.

Jev takes a state and a set of questions, then returns typed answers, probabilities, and confidence. It does not generate a paragraph; it turns the small decisions inside agents and workflows into values your code can use.

ONE REQUEST · PARALLEL ANSWERS

jev-latest

state

Three deploys failed. Production is returning 500s.

noul

Should this go to a person now?

confidence

0.97

Needs a person

state

in

questions

parallel

output

typed

70–500ms

reported end-to-end range

3

Choice · Score · Noul

Typed

define the shape first

WHAT JEV ACTUALLY IS

Give the hardest if-statements a model that can say how unsure it is.

Traditional LLMs are excellent at generating language and can be prompted to return JSON. In a real automation loop, though, every generated token, parser, and schema check adds latency and another place for failure.

Jev takes a different path: state becomes the input, questions become the program interface, and answers stay inside the shape you defined. Every answer includes probability and confidence, so your code can decide when to automate and when to ask a person.

MACHINE-NATIVE INTELLIGENCE

A model optimized for software, not for the chat window.

TypeSafe describes Jev as the first System One model: a model class built around machine-to-machine interaction. The goal is not to make every answer pleasant to read; it is to make narrow decisions observable, testable, and useful inside a larger system.

Optimize for human preference

Chat-first models

  • Write useful, natural language responses
  • A sequence of messages and generated tokens
  • Flexible strings that can express almost anything
  • Preference, helpfulness, and verifiable outcomes

Optimize for calibrated decisions

System One models

  • Make one focused decision at a time
  • A state plus typed questions
  • Constrained values and probability distributions
  • Accuracy, calibration, speed, and consistency

TWO MODELS, TWO MODES

Jev sits beside an LLM, not in place of one.

Keep open-ended reasoning and generation with an LLM. Move routing, guardrails, scoring, and triage to a model built for fast decisions software can consume.

Answers for people

Traditional LLM

  • 01Sequential, token-by-token sampling
  • 02Strings or JSON that still needs parsing
  • 03Confidence usually needs to be prompted
  • 04Chat, writing, open-ended reasoning

Decisions for software

System One · Jev

  • Multiple questions answered in parallel
  • Typed values defined in advance
  • Probability and confidence with every result
  • Routing, guardrails, scoring, triage

THREE PRIMITIVES

Shape the question so code can understand it.

One state can carry multiple questions. You define the boundary; Jev makes the decision inside it.

Choice

Pick from a set

For intent classification, model routing, and ticket triage. Returns probabilities for each option and an overall confidence.

route = [fast, powerful]

Score

Rate across levels

For risk, urgency, and quality. Returns a score, the distribution across levels, and confidence.

urgency = [low, medium, high]

Noul

Answer a yes/no question

For safety checks and escalation. Returns the probability that a statement is true—simple, but ready to drive a branch.

needs_human = true?

ASK ONE GOOD QUESTION

The best Jev question feels like a quick expert gut check.

A narrow question gives the model a clear contract. A request like “analyze this ticket and decide what to do” mixes several judgments together. Split the dimensions, then compose them in code.

Too broad

Rate this startup pitch.

This hides several independent ideas: market size, technical feasibility, differentiation, and perhaps execution risk. One score makes it difficult to inspect or change the weighting.

Atomic questions

Ask about the parts.

Score market size, technical feasibility, and differentiation separately. Your code can weight those answers, set different thresholds, and explain which signal drove the final route.

When priorities change, change a coefficient or a branch in your code instead of rewriting one giant prompt.

PUT IT BACK IN YOUR SYSTEM

From state to action, keep only the judgment you actually need.

Jev does not own your business logic. It answers bounded questions; queues, permissions, retries, and final actions remain yours.

01

Input state

A ticket, message, JSON object, or the structured context your agent sees.

02

Define questions

Write the decision you want to automate with choice, score, or noul.

03

Get probabilities

Several questions return in parallel, with structure and probability ready for code.

04

Take action

Route, block, queue, or ask a human when the signal is uncertain.

CONFIDENCE IS A CONTROL SIGNAL

The answer is only half the decision. The other half is how much to trust it.

Choice and Score return the full probability distribution plus a confidence value. A concentrated distribution suggests a clear winner; a flat distribution says the state, the question, or the options need another look. Noul returns the yes probability directly and has no separate confidence field.

HIGH

Act automatically

The signal is clear and the action is reversible or low risk.

route · show · continue

MEDIUM

Proceed with a check

Keep the workflow moving, but ask for confirmation, gather more context, or flag the case for review.

confirm · enrich · review

LOW

Do not guess

Let the model’s uncertainty change the system behavior: ask a human, request clarification, or fall back to another path.

escalate · clarify · fallback

There is no universal threshold. A read-only action can tolerate a lower threshold than a transfer, deletion, or other consequential action. Start conservatively and tune on your own labelled data.

WHERE IT FITS

The small decisions software makes thousands of times a day.

Model routing

Use a lightweight model for simple tasks and reserve expensive reasoning for requests that need it.

Tool guardrails

Check a risky action before execution. Pause on deletion, payment, or other high-impact operations.

Ticket triage

Classify intent, urgency, and human handoff together; route the queue by signal, not keyword piles.

Real-time apps

When a product cannot wait for a long answer, use a fast, predictable judgment to drive the next screen.

COMPOSABLE PATTERNS

The model stays small. The system becomes more capable.

TypeSafe’s documentation frames Jev as a set of composable decisions. These patterns turn the primitives into reusable architecture without hiding the final policy in a prompt.

01

Speculative fan-out

Ask the questions your code might need in one request, including questions that only matter for some inputs. Ignore unused answers instead of paying for a second round trip.

02

Confidence-gated routing

Use the selected option and its confidence as two separate signals. A low-confidence route can fall back to a stronger model or a human.

03

Composite scoring

Combine independent Score questions such as severity, frustration, and completeness with explicit weights owned by your application.

04

Intent routing

Classify the request into a known handler before invoking the expensive or specialized workflow that follows.

ONE REQUEST, MANY DECISIONS

Put the schema in the request. Keep the result in your system.

Jev starts with state and questions. Questions can run in parallel, and the shape of every answer is defined before the call.

REQUEST01
{
  "model": "jev-latest",
  "state": "Deploy failed twice; prod is returning 500s.",
  "questions": {
    "urgent": {
      "type": "noul",
      "instructions": "Needs attention now?"
    },
    "route": {
      "type": "choice",
      "options": ["fast", "powerful"]
    }
  }
}
RESPONSE02
{
  "urgent": {
    "noul": 0.999
  },
  "route": {
    "choice": "powerful",
    "probabilities": {
      "fast": 0.08,
      "powerful": 0.92
    }
  }
}

INSIDE THE AGENT LOOP

Put a fast decision layer around the model that writes.

LangChain exposes Jev through TypeSafeClassifier, so a Jev decision can live in a node, a middleware hook, or a tool wrapper. Let Jev decide whether a request needs a fast model, a more capable model, or a human review before the main agent spends tokens.

Read the LangChain guide
LANGCHAIN
from langchain_typesafe import Noul, TypeSafeClassifier

classifier = TypeSafeClassifier()
response = classifier.invoke({
    "state": "The deploy failed twice and prod is returning 500s.",
    "questions": {
        "urgent": Noul(
            instructions="Does this need attention now?"
        ),
    },
})

urgency = response.nouls["urgent"].noul

The same pattern can guard tool calls: inspect the intended action before execution and block or pause risky calls.

WHY SYSTEM ONE

An interface that fits automation better.

Calibration, not confidence theater

TypeSafe calls its training direction RLCD: Reinforcement Learning for Calibrated Decisions, focused on probabilities that express uncertainty.

Parallel, not queued generation

Several questions in one request can be evaluated in parallel instead of starting a complete conversation for every small decision.

Typed, not guessed after the fact

Possible answers are defined up front, so software consumes the result without guessing what a paragraph meant.

KNOW WHERE IT STOPS

A decision model is powerful because it does less.

Jev is a complement to a generative model and to deterministic code. It is a good fit when the output space is known and the question can be made specific.

  • Need a paragraph, code, explanation, or an open-ended plan? Use a generative model.

  • Need long, multi-step reasoning? Break out the atomic signals first, then let your application or a reasoning model compose them.

  • Cannot define the possible options or what each level means? Clarify the contract before calling Jev.

  • Making a consequential decision? Validate accuracy and calibration on your own data, and keep a human or deterministic policy in the loop.

FROM PLAYGROUND TO PRODUCTION

One endpoint, one server-side key, a small surface area.

The public quick start uses a single System One endpoint. Start in the Playground, move the request behind your backend, then use the official SDK or plain HTTP once the question contract is stable.

ENDPOINT

POST https://api.typesafe.ai/v1/systemone

KEY HANDLING

Keep API keys in environment variables and proxy browser requests through your own server.

PYTHON SDK

pip install typesafe-sdk

START WITH A REAL QUESTION

Give your agent one less thing to guess.

Bring a real state, define a choice, score, or noul, and inspect the result in the Playground before deciding where it belongs.