Choose one question or three different questions about the same context. Try a real scenario, edit the inputs, then inspect each typed answer.
Explore one answer type at a time. Each run uses one daily attempt.
Run a case to see one typed answer per question. Choice picks an option, Score places it on a scale, and Probability returns a value from 0 to 1.
The form above puts your text under state.input. Try one question or three named questions against that same state. Look for each result under answers.<question_id>.
The request defines the questions and answer formats. The response returns a separate result for each question. These are the fields you can inspect in the request preview and model response above.
modelstatequestions<question_id>typeinstructionscriteriaanswerschoiceprobabilitiesconfidencescorelegendnoulThe questions in a multiple-question case share one state but are evaluated independently. One answer is not fed into another. This Playground caps a run at three questions and counts it as one daily attempt. A correctly shaped answer can still be wrong, so test representative cases before automating an action.
Use an OpenRouter key with the version pinned in this Playground. Check the model page for current pricing and availability.
TypeSafe has its own endpoint, model IDs, and API keys. Its quickstart uses jev-latest and shows how to ask multiple questions in one request.
Output tokens are free.
Eligible accounts may receive promotional credits. TypeSafe does not publish a fixed amount on its public pages; check the console for your balance and terms.
An agent harness controls the model and tool loop. At a decision point, Jev reads a message, proposed tool call, or agent state and returns a typed answer. Your code decides the next step; the chat model still handles open-ended work.
ModelRouterMiddlewarebefore_agent → wrap_model_callClassify the latest user message before the run, choose a model, and keep that choice for subsequent model calls. The choice, probabilities, and confidence remain available in agent state.
AutoModeMiddlewarewrap_tool_callCheck a selected tool call before it executes. If Jev flags risk or insufficient authorization, the middleware blocks it and returns an error ToolMessage. It does not request human approval.
TypeSafeClassifierChoice · Score · NoulAdd your own checkpoint with custom middleware: pass agent messages or state to TypeSafeClassifier, ask Choice, Score, or Noul questions, then branch in code or send uncertain cases for review.
# pip install "langchain-typesafe[experimental]" langchain-openai
# Set TYPESAFE_API_KEY and OPENAI_API_KEY on your server.
from langchain.agents import create_agent
from langchain_typesafe.experimental.middleware import (
ModelChoice, ModelRouterMiddleware,
)
router = ModelRouterMiddleware(
choices={
"fast": ModelChoice(
model="openai:gpt-5.6-terra",
criteria="Simple lookups and extraction.",
),
"powerful": ModelChoice(
model="openai:gpt-6-astra",
criteria="Complex reasoning and high-stakes decisions.",
),
},
instructions="Choose the least costly model that can do the task safely.",
)
agent = create_agent("openai:gpt-5.6-terra", middleware=[router])
result = agent.invoke({"messages": [{"role": "user", "content": "Summarize this log."}]})
print(result["model_route"].choice)// Run on your server. Keep the API key out of browser code.
const apiKey = process.env.OPENROUTER_API_KEY;
if (!apiKey) throw new Error("Set OPENROUTER_API_KEY");
const response = await fetch("https://openrouter.ai/api/alpha/decisions", {
method: "POST",
headers: {
Authorization: "Bearer " + apiKey,
"Content-Type": "application/json"
},
body: JSON.stringify({
model: "typesafe/jev-1.13",
state: {input: "A customer reports a duplicate charge."},
questions: {
team: {
type: "choice",
instructions: "Which team should handle this?",
criteria: {
billing: "Payments and refunds",
technical: "Bugs and integrations"
}
},
urgency: {
type: "score",
instructions: "How urgent is the issue?",
criteria: ["Can wait", "Needs attention today", "Immediate action"]
},
escalate: {
type: "noul",
instructions: "Should a human review this first?",
criteria: {false: "No review needed", true: "Review before acting"}
}
}
})
});
if (!response.ok) throw new Error("Decision request failed");
const {answers} = await response.json();
console.log(answers.team.choice, answers.urgency.score, answers.escalate.noul);Choose the type that matches the decision your code needs to make. In the multiple-question cases above, one request returns a separate answer for each question ID.
Explore TypeSafe question types ↗Use the chosen key to route a request. Inspect the distribution before automating a high-stakes action.
The score is a probability-weighted position and may fall between the levels you defined.
Read a value from 0 (no) to 1 (yes). Noul has no separate confidence field.
TypeSafe accepts text, JSON objects, or arrays of text as state; this simple form wraps your text under input. Jev does not accept images, audio, or video. Typed output keeps the format predictable, but an individual judgment can still be wrong. Test on your own examples and define when your code should ask for review.
Track new projects, System One community models, and practical use cases through frequent public updates or an occasional curated email.