Schleuse

How the Schleuse works

We show the path of a letter. The letter is made up. It contains a name, a company, an address, an account number and an email address.

  1. The document is read

    PDF, Word, Excel, PowerPoint and scans are turned into text on your side. This happens in a protected space without a network. Macros, hidden content and encrypted files do not get through. With scans, text recognition can make mistakes.

  2. Fixed patterns find identifiers

    An account number, an email address or a metering point follows fixed rules. A program recognises them very reliably. It does not have to guess.

  3. A name detector finds names

    People, companies and places have no fixed form. A detection model estimates from the sentence what a name is. That is a probability, not certainty.

  4. A language model reads once more

    A language model on your own equipment reads the whole text in context. It finds what the others miss. For example a person who can only be recognised by their role.

  5. The findings are merged

    What any stage finds counts. Amounts, dates and quantities take priority. They must stay unchanged in the text.

  6. The key is created

    Every name gets exactly one stand-in word. It is the same throughout the document. The key stays with you, in working memory only and only for a short time.

  7. Replace, but with meaning

    The company becomes “Energy supplier Alpha”. The customer becomes “Customer Alpha”. This keeps the text understandable for the AI. Technical terms such as CHP stay as they are.

  8. Check before sending

    Does a check still find a hit from the key? Is every amount of the original also in the result? If not, the document goes into quarantine. When in doubt, nothing is sent.

The way back

  • The cloud AI answers with stand-in words.
  • The answer passes a gate. The gate can accept or reject it.
  • On your side, the key translates the stand-in words back.
  • A round-trip test checks: out and back gives exactly the original. Otherwise the request stops.
  • Control words (canary values) show whether the AI worked carefully. They must come back unchanged.

The biggest danger: silent gaps

In one test a detector read only the beginning of long texts. Another failed on very long texts without any error message. Since then every detector receives the text in overlapping pieces. Coverage tests check that nothing is missed.

Limits

  • No detection reaches one hundred percent. A small remainder is possible.
  • The more is found, the more is also replaced unnecessarily.
  • With rare roles or sole traders, even a general term can give a person away. Then a coarser level is safer.

Status of the Schleuse

Version 0.1 Alpha. So far tested only with made-up examples. Not reviewed by third parties. No legal advice.

Would you like to try this?

We are looking for pilot partners. We test the Schleuse with made-up or cleared examples first. Only then does real text go through.

Request a pilotBack to the Schleuse

This page is also available as Markdown for AI reading systems. /en/how-it-works.md