evidence

how much does limenn catch? we measure it.

we test every version of limenn against realistic documents full of the kind of personal data our users handle, and count what gets through. 120 of the 270 test documents are a set we never tune on, written blind by someone who couldn't see how limenn works. for results on public benchmarks built by others, see research.

of emails, phone numbers, IBANs, AHV numbers, card numbers and postcodes caught, in both test sets

100%

of all personal values caught in the documents we tune on

99%

of all personal values caught in the blind set, which limenn has never seen

86%

personal values in outgoing requests in our end-to-end test on claude.ai

which documents

150

documents we tune on: medical letters, HR files, leases, invoices, email threads, support chats, meeting minutes

120

blind documents of 14 other kinds: patient-portal messages, home-care notes, discharge summaries, custody letters, debt-collection letters, payroll exports, court decisions and more

languages: German, French, English and Italian

6,396

personal values, each one labelled exactly, in 16 categories

the documents are made up, so no real person's data is involved, but they are built to be hard: surnames that are also ordinary words (Sommer, Winter, Blanc), names in lowercase or capitals, six date formats per language, unlabelled ID numbers, and a person called Claude. the blind set was written by a separate author who never saw limenn's detection code, so it shows how limenn does on documents it wasn't built around. identifiers are valid: AHV numbers, IBANs and card numbers pass their checksums. we also run real PDFs and sample documents through the same test.

which information

share of personal values caught in the blind set, by category · on-device checks only

emails100%

phone numbers100%

IBANs100%

AHV / national IDs100%

card numbers100%

postcodes100%

IP addresses & URLs100%

street addresses100%

dates100%

organisations90.7%

people's names90.4%

places84.6%

reference & ID numbers81.4%

health details66.2%

salaries, debts & balances46.4%

across all categories, the on-device checks catch 86% of personal values in the blind set and 99% in the documents we tune on, and 93% of what they replace really is personal data. the gap between the two sets is the honest measure of what's left to do: salaries, debts and balances mentioned in passing, and health details written casually, are the hardest. our next-generation detector (see research) brings the blind set to 91% and health details to 80%; salaries and balances still need work. we publish the gaps because we're closing them with every release.

and it doesn't slow you down

milliseconds, median, to clean a prompt or document on your device

milliseconds to clean a 10-document dossier on the fly

640

times faster when the documents are already in your vault

179×

want to see the numbers on your own documents? we're happy to run a test with you.

get in touch ↗