Open source · Personal data detection
Every phone number in your logs. Found, and proven.
Trulydone finds personal data in messy developer text, like logs, prompts and chat transcripts, and publishes how often it is right for every type. It starts with phone numbers and card numbers, where today's free tools slip most.
01 · Why
"Redacted" is a claim. Nobody checks it.
Detection tools report what they removed, never what they missed. Published tests show how much slips through, and almost none of them use the text developers actually handle.
-
31 to 54%
Share of flagged items that were really personal data, for OpenAI's free Privacy Filter model across 32 published test sets.
Uppala, arXiv 2608.02616 · Aug 2026
-
0.35 of 1
Lowest phone number score for Presidio, the best-known open-source toolkit, across three test sets. Its best was 0.54.
Albert Sikkema, Benchmarking open-source PII detection · Jun 2026
-
90 → 70%
Card numbers Presidio caught on standard text, then on unusual formats, in a 2026 robustness study.
Mind the Gap, arXiv 2609.03464 · Sep 2026
-
0
Published accuracy tests we could find on real application logs. That is where Trulydone starts.
Our review of public benchmarks · Oct 2026
02 · How it works
Checks, not guesses.
Trulydone is built to be precise first and broad later. Each part below is the design it is being built to.
Runs where your data lives
A library and command line tool that run on your own machines and in your own pipelines. Nothing is sent anywhere to be scanned.
Proves each find
A match only counts when it passes the rules for its type: the checksum every card number carries, the numbering plan of the country a phone number belongs to. Anything uncertain is reported as unsure instead of guessed.
Learns shapes, never data
When you mark a miss, Trulydone keeps only its shape, such as phone with spaces in a JSON log line. Shapes can be shared so every install gets better, without anyone's data leaving.
03 · Accuracy, in public
One number per type, every release.
Every type gets its own published score on an open test set of messy developer text, and you can run the same test on your own data. A type is labeled beta until it reaches the goal.
The goal, on both counts: at least 95% of what Trulydone flags is really personal data, and at least 95% of the personal data present gets flagged.
| Type | Where | Stage | Measured |
|---|---|---|---|
| Phone numbers | Logs, prompts, chat transcripts | First | not yet published |
| Card numbers | Logs, prompts, chat transcripts | First | not yet published |
| Email addresses | Logs, prompts, chat transcripts | Next | not yet published |
| Names and street addresses | Free text in any language | Later | not yet published |
04 · Open source
Free to run. Free to check.
A detector you cannot inspect is one more claim to trust. The parts that do the finding, and the tests that grade them, are open.
Open source, free
- The detector and its rules for every type
- The command line tool and the library
- The open test set and the scoring code
Later, for teams
- A hosted version for teams who would rather not run it
- Tuning on your own data, inside your own cloud
- Accuracy reports you can hand to an auditor