Jev is live on Zorveus

TypeSafe AI's flagship model typesafe/jev-latest is live across the Zorveus gateway, official SDKs, and connected application endpoints. You can view its latency profiles, pricing, and live cURL templates in our Model Catalog.

Founded by Diogo Almeida (co-inventor of RLHF and InstructGPT at OpenAI) alongside Erik Gafni and Sasha Sheng, TypeSafe AI built Jev around a concept they call "System One". The name borrows from Daniel Kahneman's distinction between fast, intuitive decisions (System 1) and slow, deliberate reasoning (System 2).

Standard large language models operate like System 2. They generate text token by token. When you ask them to classify a support ticket, route a payment error, or evaluate customer intent, you write prompts asking for JSON, wait for tokens to finish streaming, and handle markdown parse errors when backticks show up in the payload.

Jev is non-autoregressive. You provide an input state (raw text or JSON) and a map of questions. The model evaluates the entire question set in a single parallel forward pass, returning typed answers and calibrated confidence scores in tens of milliseconds.

# Specifications & economics

# Three decision primitives

Jev replaces prompt text generation with three typed question types:

1. Choice. Categorical classification. You define the criteria for each category key. Jev returns the selected option along with candidate probabilities and confidence scores.
2. Noul. Binary probability. The name comes from "no/yes". It returns a continuous probability between 0.0 and 1.0 representing the likelihood that a statement is true.
3. Score. Ordinal scale rating. You provide an ordered rubric. Jev returns a normalized score and level distributions.

Because the output shape is determined by the request schema, responses never fail JSON validation.

# Using typesafe/jev-latest in Zorveus

Zorveus proxies TypeSafe directly at https://api.zorveus.com/typesafe. You use your existing ZORVEUS_API_KEY for authentication. You can also pass custom tracking headers like X-Zorveus-External-User-Id to monitor per-user usage and enforce credit limits.

You can run interactive test queries directly from the Zorveus Models Catalog or select your preferred integration method from the dropdown below:

import os
import traceback
from typesafe_sdk import Choice, Noul, TypeSafeClient, TypeSafeError

USER_MESSAGE = (
    "The checkout page is broken and payments have not gone through for hours."
)

client = TypeSafeClient(
    api_key=os.environ["ZORVEUS_API_KEY"],
    base_url="https://api.zorveus.com/typesafe",
    headers={
        "X-Zorveus-External-User-Id": "demo-user-1",
    },
)

try:
    response = client.system_one(
        model="typesafe/jev-latest",
        state={"message": USER_MESSAGE},
        questions={
            "is_urgent": Noul(
                instructions="Does this message sound urgent?",
            ),
            "team": Choice(
                instructions="Which team should handle this?",
                criteria={
                    "billing": "Payments, refunds, invoices, or money issues.",
                    "technical": "Website bugs, outages, or product errors.",
                    "support": "Customer help and general questions.",
                },
            ),
        },
    )
except TypeSafeError as exc:
    traceback.print_exc()
    raise SystemExit(f"TypeSafe request failed: {exc}")

print("Urgent:", response.answers["is_urgent"].noul)
print("Team:", response.answers["team"].choice)
print(response)

curl -X POST "https://api.zorveus.com/typesafe/v1/systemone" \
  -H "Authorization: Bearer $ZORVEUS_API_KEY" \
  -H "X-Zorveus-External-User-Id: demo-user-1" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-latest",
    "state": {
      "message": "The checkout page is broken and payments have not gone through for hours."
    },
    "questions": {
      "is_urgent": {
        "type": "noul",
        "instructions": "Does this message sound urgent?"
      },
      "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "Payments, refunds, invoices, or money issues.",
          "technical": "Website bugs, outages, or product errors.",
          "support": "Customer help and general questions."
        }
      }
    }
  }'

const USER_MESSAGE =
  "The checkout page is broken and payments have not gone through for hours.";

const response = await fetch("https://api.zorveus.com/typesafe/v1/systemone", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.ZORVEUS_API_KEY}`,
    "X-Zorveus-External-User-Id": "demo-user-1",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: "typesafe/jev-latest",
    state: { message: USER_MESSAGE },
    questions: {
      is_urgent: {
        type: "noul",
        instructions: "Does this message sound urgent?",
      },
      team: {
        type: "choice",
        instructions: "Which team should handle this?",
        criteria: {
          billing: "Payments, refunds, invoices, or money issues.",
          technical: "Website bugs, outages, or product errors.",
          support: "Customer help and general questions.",
        },
      },
    },
  }),
});

if (!response.ok) {
  const errorText = await response.text();
  throw new Error(`TypeSafe request failed: ${response.status} ${errorText}`);
}

const data = await response.json();
console.log("Urgent probability:", data.answers?.is_urgent?.noul);
console.log("Team choice:", data.answers?.team?.choice);

# Why System One matters in software pipelines

Most automated agent loops do not need paragraphs of prose to take their next action. They need to know if an error is fatal, whether an input contains sensitive data, or which handler function to invoke.

Using autoregressive models for these decisions introduces real drawbacks:

• High latency. Waiting 1 to 3 seconds for a generative model to finish its thought introduces delays in automated user-facing flows.
• Format fragility. Autoregressive outputs can fail when models wrap JSON inside markdown blocks or drift across version updates.
• Excessive token spend. Paying for generation tokens on simple classification tasks adds up quickly at high query volumes.

By executing decisions in a single pass without token generation, typesafe/jev-latest delivers responses in 70 to 500 ms at $0.042 per million input tokens.

# Zorveus Edge features

When routing typesafe/jev-latest through Zorveus:

1. Unified authentication. Access TypeSafe models using your existing Zorveus API keys alongside OpenAI, Anthropic, Google, and open-weight models.
2. Hard spend caps. Set monthly and daily dollar limits on your API keys so background automation cannot exceed budget.
3. End-user attribution. Attach X-Zorveus-External-User-Id to any request to track token usage and spend down to individual users.
4. Audit logs. Inspect payloads, latencies, and status codes in real time from your Zorveus organization dashboard.

typesafe/jev-latest is available in the Zorveus Model Catalog to all Zorveus organizations starting today.