Data demos are structurally similar: a polished sample, a dashboard, a claim about coverage. They are also structurally unable to show you the thing you are buying, which is what happens in week three.
These five questions are not adversarial. They are the ones your own team will ask after the contract is signed.
1. What is one row, exactly
If the answer varies by collection, the schema is not stable. If it varies within a collection, the dataset is a pile.
2. Which fields are optional
Every card lists fields. Few list the ones that are empty a third of the time, and those are the ones that break a model quietly.
3. What did you exclude
The exclusions tell you the shape of the data more reliably than the totals. A vendor with a written exclusion list has looked at their own set.
4. How are disagreements resolved
Two labelers, one row. The process matters more than the agreement rate, because the process is what you will inherit when you ask for a change.
5. What happens when a batch is late
The answer tells you whether there is an inventory or a scramble. Ask what happened the last time it happened.
Listen for the shape of the answer
A short, specific answer is a good sign. A long, general one usually means the question has not been asked before, by anyone, including internally.
Zovda answers these five on the data card for each collection shipped through Datuva. The data research collection shows the format.
Ask us all five. We reply within one business day.
