A field guide to synthetic experiments

Human behavior.
A question at a time.

Design controlled experiments with synthetic respondents. Estimate what changes choice, compare future actions, and inspect the evidence before you act.

Design your first studyChoose a business studyStart with the API

For researchers in economics, sociology, and psychology.

Explore a choice task

What makes a job worth taking?

Compare compensation, flexibility, and commute in a job-choice experiment.

Illustrative design. No experiment is running and no responses are collected.

Attribute	Option A	Option B
Annual salary	$70,000	$80,000
Working arrangement	Three days at home	Five days in office
Commute each way	45 minutes	20 minutes

What you could learn
Estimate how modeled job preference changes within the salary and working conditions you test.

This does not estimate labor-market wages or actual job acceptance.

Build this study design

A path from question to interpretation.

1. Define the decision

   Name the chooser, the alternatives, and the result that would change your mind.

2. Design the comparison

   Select a population and vary realistic attributes without changing the question halfway through.

3. Run and keep the record

   Review cost and privacy, retain the run ID, and record the settings needed to repeat the study.

4. Compare future actions

   Use the fitted model to compare feasible scenarios. Check uncertainty and relevant human evidence before acting.

Work the way you research.

In the browser

Follow the experiment builder from a question through population, design, and results.

Use the visual workflow

In your analysis code

Use a Python workflow or look up exact request fields, defaults, responses, and errors.

Run with PythonRead the API reference

With an assistant

Connect an MCP client, review the design, and make the decision to launch explicit.

Connect through MCP

Synthetic responses.
Explicit limits.

These experiments measure model responses to a design. They are not observations of people. A useful finding still needs a justified population, a defensible comparison, and validation appropriate to the decision.

Read the research validity checklistExplore human-baseline studies

A benchmark supports the study and metric it measures. It does not certify every new population or research question.