Where Studio falls short today
Known limitations
This is the boundary of what Misata Studio actually does, found by testing our own output rather than assuming it. What Studio does guarantee is checked on every dataset and listed in its certificate. Everything below is what those checks do not cover, or what the product does not do yet. If something elsewhere on this site seems to contradict this page, this page wins, and the other page has a bug worth reporting.
Last updated 22 October 2026.
Using Studio
A dataset takes about one to two minutes, not seconds
Making and checking the rows takes seconds. The rest goes to the models: writing content for your business (the catalog, wording, people and places, seasons and research) and an expert pass that reads the finished data. The run screen shows each step as it happens. You can cancel a run (the credits come back), or leave the page: a reload or a return picks the build up where it is.
A build on your own model key stops if our server restarts
A build runs on our server, not in your browser, so closing the tab or losing the connection does not stop it. When our server restarts (a deploy), a build on the platform's model is taken up again under the same id, with the model answers it already had. A build on your own key is not, since the key is never stored: it is marked as not finished and you get an email saying so; building it again with the same words usually works.
Generated files are temporary; the recipe is what lasts
Files are kept for a limited time (up to 14 days where archiving is on, and until the server restarts where it is not). The recipe, your request and the settings that rebuild the data, is kept with your account, so an expired dataset can be regenerated with one click.
The no-sign-up demos run on the earlier engine
The no-sign-up demos are paused while Studio is invite-only. The public try page, the predictive maintenance generator and the warehouse generator still use the earlier, model-free engine. They show connected tables and integrity checks, but not Studio’s written content, web research, requirement list or certificate. Moving them to the Studio engine is planned.
Studio hands over files; only the API loads a database
In Studio you download files or copy a ready-made command (Postgres, Snowflake, BigQuery, Databricks, DuckDB, S3). The platform API loads Postgres and ClickHouse directly, over an encrypted connection to a host on the public internet; private networks are refused, and nothing is loaded into a bucket for you.
Numbers and rules
Stated rates round to whole rows
A 2% rate over 4,824 rows cannot be exactly 2.00%: 96 flagged rows is 1.99% and 97 is 2.01%. Studio uses the closest whole number of rows, and the certificate shows the figure actually reached.
A drawn curve can set totals, not rates
You can draw monthly totals (revenue, volume) and each month lands on your figure. A rate that changes over time, such as churn rising through the year, cannot be drawn yet. A single rate you state in words is applied exactly.
Correlations exist only where something puts them there
Studio does not learn from real data. Relationships between columns come from a tested archetype's mechanics, from rules you state, or from the realism pass (a fee that depends on the procedure, a category that follows a number). Subtler real-world dependence is absent unless one of those sets it.
Anything that cannot be met is reported, not bent
If two things you asked for cannot both hold, or a figure cannot be applied, the requirement is listed as not met or not applied. Studio does not quietly change what you asked for, which also means some requests come back with an honest gap.
Content, worlds and research
Text is assembled, and repeats in large tables
Reviews, notes and descriptions are built from lines a model writes for your business, chosen to match each row (a low rating reads as a complaint). In a large table the same lines come back, and long passages have a rhythm a careful reader can spot.
Domains without an archetype are less consistent
Ecommerce, SaaS and sales, fintech, healthcare, clinical trials, education, AML and machine telemetry run on tested archetypes. Anything else is designed from your description, which is slower and varies more between runs. You see and can change the tables before any data is made.
Invented worlds borrow real mechanics
A shop on Mars in 2050 gets its own peoples, places, money and wording, but underneath it runs a real archetype such as ecommerce. Column names may keep Earth units, and money may stay in dollars; where that happens, the dataset says so.
Research can bring in a fact that does not belong
Facts from the web are kept only if their quoted words appear on the source page, and every source is linked on the dataset. Whether a fact is relevant is still judged by a model, so an off-topic fact can occasionally get through. Check the sources listed, or switch research off for that dataset.
Places are checked most thoroughly for about 65 countries
Names, cities, phone numbers and postal codes are written per country and checked against each other. Phone formats are validated in depth for about 65 countries; elsewhere they are plausible but checked less strictly.
The expert-style review is a model's opinion
It reads summaries of the finished data and flags what looks wrong, and code checks its fixes before they are applied. It catches a great deal, not everything, and is not a substitute for someone who knows your domain.
Scale and repeatability
About 10 million rows per dataset; star and flat layouts to about 3 million
Large runs are generated in parts with totals still exact across them. Star-schema and single-table layouts are built in memory, so they stop at about 3 million rows.
Repeatable per engine version
The same recipe and seed rebuild the same data while the engine version and the stored model answers are unchanged. An engine upgrade or a fresh model answer can change individual values; the numbers you stated, the keys and the rules still hold.
Every entry here started as a reproduced defect or a deliberate design choice. When one is fixed it moves to the changelog. The open-source Python library has its own, separate list, since it works differently (no models, no research, runs on your machine).
The library’s limitations on GitHubFound a failure not listed here? Tell us, that is a bug in this page too.