Aktenlage · technical

How it works

The path from the sheet in the feeder to the filed PDF, the rule set behind it, and the figures measured in real use.

01

What happens between feeder and filing

Every step also runs on its own and records what it did.

Step 1

Scan

The scanner delivers untouched raw images. Its own functions for discarding, deskewing and rotating stay switched off.

Step 2

Prepare

Sort out blank pages, detect orientation, trim margins, recognise text. Every step can be repeated and undone.

Step 3

Split

Where does one letter end and the next begin? Eight signals are weighted and combined, without separator sheets in between.

Step 4

Understand

First fixed rules with check digits, then a language model on the same machine. Contradictions between the two are recorded rather than overwritten.

Step 5

File

A searchable PDF, with a readable sidecar file next to it. Folder and file name follow a freely configurable pattern.

Branch: quarantine

Anything recognised as blank is not deleted but put aside, with the measured value beside it.

Branch: review stack

Uncertain documents wait for approval. Confidence decides the order they are presented in, not whether they are presented.

When in doubt, keep it. Nothing is thrown away. The price: the odd surplus blank page and a longer review stack.

02

How the program reads

Calculate first, interpret second.

First, fixed rules

IBAN per ISO 7064, German tax ID per modulo 11,10, VAT ID, date formats, amounts, twelve kinds of reference number, postal address, contact details.

Calculated, not guessed: every find has to satisfy its check digit. A DEO2 only becomes DE02 if the checksum works out afterwards.

Then the language model

It runs on the same machine and is handed the verified finds up front. Its answer is checked against them; contradictions are recorded.

Splitting without separator sheets

Eight signals: page numbering, reference number, change of sender, letterhead, salutation, change of format, change of date, short closing page. Plus the test of whether a document has as many pages as it claims.

Blank pages without losses

Counting dark pixels fails on shadows, punch holes and show-through. Aktenlage subtracts from the back side whatever the front side explains. What remains is content of its own.

03

The archive stays readable without the program

No proprietary file format, no tie to the vendor.

  • The file system is the truth

    Next to every PDF sits a readable text file with all the fields. The database is only an index and can be rebuilt from it.

  • Moving without tools

    Copy the folder, rebuild the index, done. Anyone who stops using it keeps searchable PDFs with plain text alongside.

  • Nothing is deleted

    Blank pages go to quarantine and can be fetched back. Duplicates are reported, not discarded.

  • Approval only by the user

    No document counts as checked automatically. Every file carries its checksum.

04

What leaves the machine

The documents never. Otherwise one recurring connection and a one-time download at the push of a button.

OperationWhat is transmitted
Scanning, recognising, understanding, filingnothing
Language model
talks to a service on the same machine (127.0.0.1)
nothing, as long as the preset address stands. It can be changed; anyone who enters a remote machine there sends the document text to it.
Licence checknothing
Searching the archivenothing
Checking whether a newer release exists
at most once a week, can be switched off
a request for a public file on timobritz.de. No identifier, no licence number, no counting. Nothing is downloaded.
Loading the language model once
only at the push of a button, about 5 GB
the model is fetched through the bundled Ollama service from Ollama's public model registry. The provider sees the IP address and the model name; no documents, no licence data.

The release file is signed. If it fails the check, the program shows neither a message nor a link. A forged update notice therefore does not work.

For law firms and medical practices: since no data goes to third parties, no data processing agreement is required.

05

Measured, not estimated

From a real run on an ordinary desktop machine. The figures depend on the originals and the hardware and are not guaranteed.

1.3 s
per page for text recognition
confidence 92 to 93
4.7 s
per document for the language model
entirely on the graphics card
7 / 7
classifications correct
category and document type
740 KB
per page in the archive
colour, 300 dpi, searchable
571
tests, all passing
re-run for every release

Testing concentrates on the places where an archive takes quiet damage: blank pages, Windows file names, sender assignment, rebuilding the index, single flipped bits.

06

State of development

Eleven releases since 1 September 2026. All 1.x are included in the price.

1.0 to 1.2

The chain runs end to end

Scanning, preparing, splitting, understanding, filing, searching. All settings in the program, statistics, backup, duplicate detection.

1.3 and 1.4

Searching and maintaining

Search with preview and filters. Re-evaluation of the whole archive: when recognition improves, the archive runs through again.

1.4.2 to 1.4.4

What actual use turned up

Backups could not be read back in, nothing came back out of quarantine, document dates could lie in the future. All fixed.

1.5.0

No second installation step

The service for the language model now ships with it. And without the model, every document used to end up under „Sonstiges“.

Next up

Merging and splitting by hand

The biggest known gap.

07

Back to the program

Price, screenshots, limits and download are on the product page.