How we verify your data
Every dataset Misata Studio builds comes with a certificate. It lists what you asked for, whether each requirement was met, and the checks that were run on the finished tables. This page explains those checks and, just as importantly, what they do not prove.
Last updated 21 September 2026.
The principle
Models help design the data, but code decides whether it passes. The checks read the finished tables, not the intention behind them, so a mistake in the design cannot mark itself as correct. When something fails, the certificate says so and shows which requirement.
What is checked
Your numbers, stated against actual
Every total, rate, count and range you asked for is listed on the certificate with the figure you stated and the figure in the data. A requirement is marked met only when they agree.
- Totals by month or year, including a curve you drew
- Rates and shares, such as a 7 percent denial rate
- Counts, such as 3,000 patients
- Bounds, such as no fee above a stated maximum
Structure
The tables have to fit together the way a real database does.
- Every foreign key points to a row that exists (no orphans)
- Primary keys and other unique columns are unique
- A child row exists only when its parent allows it, such as a payment only for a completed visit
- Formulas hold, such as a line total equal to price times quantity
Time makes sense
Dates are checked across tables, not just within one.
- A visit is never before the patient registered, a payment never before its visit
- Events fall inside the period you asked for
- Rows that could not exist yet are not created
Realism
Automatic checks for the tells that make data look fake.
- Values inside believable ranges for what the column means
- Names, cities, phones and postal codes that agree with each other and the place
- No placeholder text or vocabulary that leaked from another domain
- No repeated patterns that only a generator produces
An expert-style review
A language model reads a summary of the finished data as a domain practitioner would, and flags what looks wrong: a fee that is too small, a value that cannot exist in that state. Its fixes are checked by code before they are applied, the data is rebuilt once, and the rebuilt version is kept only if it still passes every check above.
- It sees ranges and category shares, not your rows
- Its suggestions are limited to a fixed set of edits code can verify
- Anything it cannot express safely is listed as a note for you
Check it yourself
You do not have to take the certificate on trust. Open the dataset, go to Explore, and run your own SQL against the tables in your browser: add up a month, count the orphans, compare two dates. Each requirement on the certificate names the column it was measured on, so you can reproduce the figure.
What it does not prove
- The data is invented. It matches the world you described, not a real population, so it is not a forecast and not evidence about real people.
- The expert-style review is a model's opinion. It catches a great deal, not everything, and it is not a substitute for someone who knows your domain looking at it.
- Free text, such as review comments, reads naturally but has patterns a careful reader can spot.
- Worlds with no data behind them (a colony in 2050) are written with cited research where it exists; the sources are listed on the dataset, and the rest is stated assumption.
More detail is on the known limitations page.