What the score means, and what happens if we got it wrong.
Three questions worth answering before you spend anything: how a dataset earns its number, what you can do if it turns out not to match, and who carries the risk if a seller was wrong about where the data came from.
What the quality score measures
Every dataset is profiled automatically before it can be listed. Five dimensions, weighted, combined into one number from 0 to 100. Here is exactly what goes into it.
Completeness
30 %How many values are actually there. Every column is measured for missing values, so a dataset that looks full but is 40 % empty cannot hide it.
Consistency
20 %Duplicate rows and columns that mix types, like numbers stored as text. These are the errors that break a pipeline three weeks after you buy.
Schema quality
20 %Whether columns are named and documented well enough to be usable by someone who did not build the file.
Size adequacy
15 %Whether there are enough rows for the dataset to be worth anything for analysis or training.
RGPD readiness
15 %Result of the automatic personal data scan, cross checked against what the seller declared. Undeclared personal data pulls the score down hard.
The five scores are combined into a single number from 0 to 100, which sets the label shown on the listing:
This is an automated check on the shape of the data, and we would rather say so plainly. It tells you whether a file is complete, coherent, documented, big enough and clean of personal data. It cannot tell you whether the numbers in it are true, or whether the seller really collected them the way they say. That is what the escrow window and the dispute process below are for.
If a dataset is not what it claimed
You can contest a purchase when the file does not match its description, the announced metadata is plainly wrong, it is unusable, or it contains personal data that was never declared.
You raise it
From your dashboard, within 14 days of delivery. You say what does not match the description.
The money stops
Funds are already in escrow. They stay there while we look, so nothing has reached the seller yet.
We look at both sides
Your account, the seller response, and the verification report generated when the dataset was listed.
We decide, in writing
Within 10 working days, with the reasoning. If you are right, you are refunded in full to your card. Not a credit.
Who decides. The datrust team. We are the intermediary, so we are the ones who arbitrate, and we tell you why. A seller who gets this wrong repeatedly loses their listings and their account.
Timings and the full procedure are in our conditions générales, article 8.
Licence and liability
The question nobody likes to put in writing: if a seller lied about provenance and you trained a model on the data, who is exposed?
The seller carries it
Every seller warrants, per listing, that they hold the rights to sell the data, that it was collected lawfully, and that the provenance and licence they declared are accurate. If that warranty is false, the liability is theirs, not yours and not ours.
What you get from us
A full refund under the process above, the verification report and PII scan we ran at listing time, and the seller declarations on record. Those are the documents you would need to show you carried out reasonable diligence. Our own liability is capped at what you paid for the transaction, and we would rather you know that up front than discover it in a clause.
What you should still check
The licence on the listing tells you what you are allowed to do with the file. Training a model is not automatically covered by every licence, so if that is your use case, read that field before buying and ask us if it is ambiguous. Redistribution and resale are not permitted unless the licence explicitly says so.
Full wording in our conditions générales, articles 4, 5 and 9. Written to be read, not to be survived.
Try it on your own file first
The same scoring engine, free, on any file you have. No account, and your file is analysed in memory then destroyed.