Quickstart

Take a dataset through a first analysis, inspect the results, and save the work. This walkthrough requires access to a working development build; public installers are not yet available.

Before you start. Launch your build and confirm it can reach its configured model service. Read availability and setupif you do not have a build. Use data you are allowed to send as analysis context; see Security & data flow.

Step 1: Open Mizura

Launch the app. The workspace asks What do you want to investigate? No session exists yet. One is created the first time you send a message. That is the whole setup.

Step 2: Load a dataset

Drop a dataset file onto the workspace, or use the dataset panel on the left. CSV, Excel, Parquet, JSON, Stata .dta, and more. The full list is in Your data. Loading gives you an immediate preview: row and column counts, the first rows, and summary statistics for the numeric columns.

An advertising dataset loaded in Mizura, with spend and sales columns previewed beside the research question — A loaded dataset, inspected alongside the research question.

Tip. Look at the preview before you ask anything. Check column names, variable types, missing values. Importing a file successfully does not mean the dataset is ready for inference. Your data lists what to look for.

Try a small example

Download first-analysis.csvand load it as a dataset. It contains 24 synthetic observations withobservation, group, spend, andsales columns, with no missing cells. These values are for learning the workflow and support no real-world research conclusion.

Inspect this synthetic dataset. Confirm the row count, column types, and missing values. Plot sales against spend and describe the observed association. Print the descriptive statistics you use. Do not claim causation or change the dataset.

Success means you can find a data summary, the requested plot, and the Python that produced the numbers. Confirm the preview has 24 rows and four columns. If the agent reports another size, check the file and active dataset before proceeding. Generated wording and plot styling can vary between runs.

For a deterministic check, download the sandbox Python exampleand inspect it before asking Mizura to run that exact code with the sample dataset. It creates a descriptive table and scatter plot and prints 24 rows, mean spend 47.5, and mean sales 49.5. This script uses Mizura's injected data, save_table, andsave_figure helpers; it is not a standalone Python program. Generated prose can vary, but these computed values should match.

Step 3: Ask a question

Plain language, English or Arabic. Say what you want and what you want back:

Does monthly sunspot activity follow a seasonal pattern?
Fit an appropriate time-series model and report the evidence.

Mention the outcome and predictors. Mention the population or time window. Say whether you want a test, an estimate, or a plot. Name any decision you have already made, like a significance level or a reporting style. Attachments go in with the + button, by dragging, or by pasting. Five per message at most. Codebooks, questionnaires, and PDFs count as context, not datasets.

While the agent works you see the thinking trace, the tool calls, the live Python in the sandbox panel, and the answer streaming in. Workflow & plans covers the workspace in full.

Step 4: Review the plan, then press Go

For substantial work, or when you ask to review the approach first, the agent can propose a plan. It puts the plan in the Plan tab and stops. Nothing in that planned analysis runs until you press Go.

Read the steps before you approve them. Three questions matter. Does the method fit the question and the data? Can the assumptions it names actually be checked in this dataset? Is something missing, a variable, a subgroup, a sensitivity check?

To change the plan, use Review and revise and say what should change. Revisions do not approve anything. You still press Go when the steps look right. An ordinary chat message never approves a plan either.

A proposed subgroup stability plan in the Plan panel, awaiting approval with Review and revise and Go buttons visible — A proposed plan, ready for review before anything runs.

Step 5: Review the result and export it

Results land in the Outputs tab. Figures, tables, scripts, reports. Each one carries the code that produced it. Open the script. Read what actually ran. Open the figure or table and inspect it.

Check two things on every result. First, the numbers. Every statistic in the answer has to trace back to executed code. A flagged answer is unfinished, so ask the agent to compute what is missing. Evidence & literature explains the check. Second, the assumptions. The method card names diagnostics that should have been checked. Confirm they appear in the run, and that they support the conclusion.

Worth keeping? Ask for a formal report, or export the session as a notebook, a Quarto document, or a replication package. Reports & exports covers all three.

Residuals diagnostics figures open in the Mizura output viewer, zoomed to 150 percent — A generated figure, reviewed beside the question that produced it.

Next. Before your second session, read Your data. It covers importing from GitHub, Google Drive, and Zenodo, plus the size limits on large datasets.