prevalenceFAQ
Straight answers about the number and how we stand behind it.
The plain-language version of what we measure, why the error bar is honest, and what touches your data. If your question is not here, tell us about your dataset and we will answer it directly.
Frequently asked
One number, and the reasons you can trust it.
What number do you give me?
One number: the fraction of your documents that contain personal information, with a 95% margin of error. For example, "68%, and we are 95% confident the true figure is between 63% and 73%." It is one number for the whole dataset, not a pile of alerts to triage. A narrower margin means a more precise answer, and narrowing it without overstating what the sample supports is the point of the work.
How is your error bar honest?
Two ways. First, we measured it against reality: on public datasets where the true answer was already known, our margin of error contained the true figure every time. Second, the number is corrected in both directions. A detector that over-flags pulls the honest number down; a detector that misses things pulls it up. We are not trying to hand you a low number or a high one. We give you the accurate one, with a written method anyone can re-check.
Does it work with the scanner I already use?
Yes. The method runs on top of your existing detector, including Microsoft Presidio and common DLP tools, as well as ours. We measure how often your scanner is right and wrong against a small answer key your reviewers label, then correct its count for that error. A sharper scanner gives you a tighter margin of error; a rougher one still lands on the right answer, just with a wider margin. You are not locked into our detector.
Do you see my data?
Not unless you want us to. The audit can run entirely on your own computers, with nothing sent out. In our own testing, a local model running on a laptop matched or beat a much larger hosted model on most kinds of text, so staying inside your perimeter does not cost you the quality of the result. Your reviewers label the small answer key on your systems too.
Why isn't a scanner enough on its own?
A scanner flags matches one document at a time. Across a dataset too big to read, that is millions of flags nobody can review, and a raw count of matches is not something you can defend to a regulator, an auditor, or a buyer's diligence team. It is also usually too high, because detectors are tuned to over-flag rather than miss, so you end up redacting and reviewing data that was never personal. We check a representative sample instead, correct for how often the detector is wrong, and report one rate for the whole dataset with a margin of error you can stand behind.
Has this run on my kind of data yet?
The method is tested on public datasets where the true answer was known, across legal, clinical, and general text in two languages, and the margin of error landed on target each time. It has not yet run on a live customer dataset. The first project is exactly that check: we size the sample with you, agree with your reviewers on what counts as personal information, and deliver a number validated on your data. The margin of error covers sampling; it does not settle what counts, which we pin down with you up front.
Still have a question?
Tell us about your data and what you are worried about. We come back with a sampling plan, a labeling budget, and a fixed scope.
All demo data is synthetic. We never ship real data out to audit your data.