Easy-language drafts of Swiss official texts

Before and after

A sentence from an official letter, in "Behördensprache" and with three negations. A reader with low reading skills has only one question: do I get my money back? The easy-language version answers it in three short sentences.

How it works

The language model does everything that needs language understanding. Plain code does everything that can be counted.

1. Understand (model)

   The document becomes a fact sheet: the reader's core questions, every fact with its exact value, hard words. Code adds every date, amount and contact the model missed.

2. Write (model)

   A draft in easy language, every sentence linked to the source sentences it comes from, every omission explained.

3. Check (code)

   Deterministic rules from the published Swiss guidelines and an exact check of every fact value in both directions: no value lost, no value invented.

4. Judge (model)

   Meaning only: distorted facts, invented statements, metaphors, unclear references.

5. Repair (model)

   Only the flagged sentences, until no blocking finding is left; then one sentence for each must-keep value still missing.

6. Review (person)

   Source and draft side by side, provenance, fact checklist; approve, edit or reject each block, export with an audit trail.

Measured, not promised

The full pipeline against a strong single prompt (a published Leichte Sprache prompt from a Swiss public administration) on the same model: 20 held-out pairs of real German texts published by Swiss cantons and federal offices with their easy-language versions, Apertus on the CSCS endpoint, 12 October 2026.

All numbers and the method: evaluation.md.

Roadmap

What has shipped since work began in October 2026, and what comes next. Ideas and requests are welcome as GitHub issues. If your administration needs something from the list sooner, or something that is not on it, tell us.

1. 0.5 and 0.6 October 2026

   German and French

   Leichte Sprache, Einfache Sprache and FALC: the fact sheet, the rule checks from the published Swiss guidelines, the judge, the repair loop and the review page; exports as Markdown, HTML and Word with an audit trail. A decision mode for KESB and court decisions keeps the operative part and the appeal instructions. The evaluation on 20 held-out pairs (results above).

2. 0.7 October 2026

   Any model, one command

   On PyPI: uvx leichtfacile serve downloads and starts the app. It works with any OpenAI-compatible endpoint, with its settings in one file and ready-made presets for Apertus and the large providers.

3. 0.8 October 2026

   Italian

   Lingua facile with 28 rules of its own, a tagger built from the Morph-it! lexicon, vocabulary, prompts, interface and an evaluation on Public AI. The app's model page sets up the endpoint and its key.

4. Next

   German for Germany and Austria

   The rules of the Netzwerk Leichte Sprache and DIN SPEC 33429, each country's spelling with ß, its numbers, dates and phone numbers, and prompts and vocabulary with the words of its authorities. The Swiss variant stays unchanged.

5. Planned

   Letters someone else wrote

   The person who received a letter, or someone who helps them, uploads it. Code lists every deadline, amount and contact in the letter's own wording, a fixed notice says that only the letter counts, and the explanation gives no advice.

6. Planned

   Decisions: detection and long reasoning

   The app recognises the type of document instead of asking for the mode, and a further mode condenses long legal reasoning and lists every omission.

7. Planned

   More kinds of documents

   Discharge letters, lease terminations, penal orders, payment orders, school decisions and information for study participants, each with the periods and values its readers must not miss (use cases).

8. Research

   Scanned letters and photos

   Apertus 1.5 reads images, but on whole pages it turned IBANs and periods into other plausible numbers. Sent in strips at the scan's resolution, it kept every digit of two test letters. For now, a PDF without a text layer gets a clear error message.

Authors

Alexander Bobrovsky

25+ years in IT, 15+ of them in product and program management, with a focus on system design and artificial intelligence. Designed and delivered systems with five-nines availability that process hundreds of millions of events per day.

github.com/bobrovsky420

Ivan Kuzmych

15+ years of hands-on software development in Java, Python and JavaScript/TypeScript: production systems built and shipped end to end, DevOps and CI/CD included.

github.com/ikzmch

Support for public administrations

Leicht·Facile is free and open source, and you can run it on your own. If you work for a public administration and would like help, we, both authors, support you on a voluntary basis and free of charge:

• Putting it into operation

  Installation on your servers or in your cloud, connection to your model endpoint (Apertus on your own hardware, the CSCS inference API, Public AI or another provider), hardware sizing, and air-gapped installation, where the app, the model and all data run in a network with no connection to the internet.

• Adjustment to your texts

  Your terms and vocabulary, the facts your letters must never lose, your house rules for easy language, your types of documents.

• Integration

  Connection to your document management or case management system.

• Evaluation with your own texts

  We measure the drafts on your letters and their easy-language versions, with the method behind the results above, so you see what the app does before you start.

Tell us briefly who you are, which texts and languages you have in mind and where the app should run, and we will get back to you.

Or write directly to leichtfacile@gmx.ch.