Mündigkeit

For learners themselves, and for teachers in integration and vocational language courses, migration counsellors, social workers, staff in local authorities and libraries, journalists, and researchers in migration, education, and law.

The structural access gap

The German citizenship test is made up of 300 nationwide questions and 10 state-specific questions — 310 in total, of which 33 are drawn for the actual exam (three of the ten state questions plus 30 nationwide). These texts and answers are official works under § 5 UrhG and are therefore exempt from copyright. They may be reproduced, translated, and redistributed without restriction. In legal terms, the learning material is public property.

In practical terms, however, it is locked behind language:

• The official BAMF online practice portal offers the questions in German only.
• Commercial preparation materials are overwhelmingly in German, occasionally in English; multilingual offerings with real depth are missing.
• Anyone who does not qualify for a state-funded integration course — EU citizens, long-term residents from non-EU countries, family members joining through reunification — is left to self-study or paid third-party offerings.

Copyright opens the material; language closes it again. Mündigkeit answers this structural access gap with 50 translations, bilingual display, and original audio.

StAG § 10 (1) — No. 6 versus No. 7: when the means becomes the precondition

The German Nationality Act lists, among other naturalisation requirements in § 10 (1), two conditions side by side:

• No. 6 — adequate knowledge of the German language (in practice CEFR B1) — language as a means.
• No. 7 — knowledge of the legal and social order and of conditions of life in Germany, tested through the citizenship test — civic literacy as the end.

On the page, the two requirements stand on equal footing. In implementation, the BAMF makes civic learning material available only in German. The operational consequence: anyone who has not yet met No. 6 cannot learn No. 7. The means becomes the precondition of the end; civic learning, which could stand on its own, is forced to wait for language acquisition instead of running alongside it.

This inversion of ends and means overlooks a basic point of pedagogy: civic literacy — knowledge of the constitution, the separation of powers, historical responsibility, and social institutions — takes hold most reliably in the language in which a learner has built their world of meaning. The point is meaning, not vocabulary. Mündigkeit therefore separates the two requirements again. Civic content is available in the learner's first language; the German original and the German audio carry, in parallel, the gradual acquisition of the specialised vocabulary needed in the exam itself and in dealings with authorities.

EU citizens in the gap: Article 21 TFEU and the Freedom of Movement Act

Under Article 21 TFEU and the Freedom of Movement Act for EU Citizens (FreizügG/EU), EU citizens have the right to live, work, draw social benefits, and bring family members to Germany — without a residence permit, without proof of language, without an integration-course obligation. From this comprehensive freedom of movement, German migration law concludes that EU citizens are not required to attend the integration course and, as a rule, do not have funded access to it. The underlying assumption: anyone who already has the right of residence by virtue of EU citizenship needs no further state support for language or civic learning.

In everyday life, this opens a gap:

• EU citizens who have lived in Germany for years or decades without ever working through the political system, the history, and the Basic Law in any structured way.
• Political participation at the federal and state level that is not available without German citizenship (Bundestag and Landtag elections).
• Daily encounters with schools, authorities, and neighbourhoods where civic vocabulary is missing.
• An institutional support landscape that explicitly excludes this group — precisely because they already have rights.

A growing number of EU citizens are also considering German naturalisation — for the federal vote, for consular protection outside Europe, for the multiple-citizenship option that has, in principle, been available since the 2024 StAG reform, for a long-standing tie to the country. They too must pass the citizenship test and meet § 10 (1) Nos. 6 and 7 StAG — and they must do so without a comparable preparation infrastructure.

Mündigkeit opens the same door for everyone in this gap: EU citizens outside the integration course, EU citizens considering naturalisation, long-term residents from non-EU countries, family members joining through reunification, and beneficiaries of international protection with other learning histories.

Civic literacy is not a private matter for people seeking naturalisation

Knowing the constitution, the history, and the institutions of a democratic society is not something that only concerns those filing a naturalisation application. It matters to:

• second- and third-generation residents born in Germany whose first language is not German,
• long-term residents with no immediate plan to naturalise,
• students, researchers, and skilled workers on time-limited residence,
• teachers and social workers working with migrant communities,
• anyone who wants to understand how the country they live in actually works.

For these readers, the 310 questions are less a test than an entry point into a democratic society. Exam preparation becomes a by-product of a civic education that can stand on its own.

What Mündigkeit offers

Three learning modes

• Practice mode — questions in whichever order you choose, one at a time. After each answer, the first-language translation, the German original, and a short legal note appear straight away.
• Full exam — the original format: 33 questions in 60 minutes, marked as passed from 17 correct answers, with a breakdown by topic area.
• State exam — the 10 questions for the chosen federal state. The initial release covers Lower Saxony; expansion to all 16 states follows in phase 2.

17 categories across 3 topic areas

The content follows the three official topic areas of the BAMF framework curriculum:

• I. Living in a democracy (9 categories) — Basic Law, electoral law, parties, federalism, the judiciary, the media, civil rights, freedom of religion, equality.
• II. History and responsibility (4 categories) — the Weimar Republic, National Socialism, the post-war period and division, reunification.
• III. People and society (4 categories) — education, social security, work, family and diversity.

The categories work as filterable tags, so learners can revise specific topics on their own.

Bilingual display

Every question shows the German original and the chosen first-language translation at the same time. The original sits at the top at full size; the translation appears below it as a scaffold for understanding. Where legal and administrative terms come up, learners can compare the two versions side by side — the kind of negotiation that research on heritage-language learners (Polinsky, Montrul, and others) describes as particularly productive for retention.

Civic Audio — the original text with Neural TTS

The German original of every question is available as audio, generated through Azure Neural TTS with the voice de-DE-KatjaNeural. The first-language translation is deliberately not narrated — the path from script to sound to meaning stays on German and gradually tunes the ear to the diction that learners will actually meet in the exam and in administrative conversations. For learners with visual impairments or dyslexia, the audio also opens a more accessible route into the material.

51 languages with qualityTier transparency Planned

The full version will cover 51 languages (German plus 50 translations) and will include all 24 official EU languages in full. Each language is assigned a four-tier quality marking:

• verified — reviewed and signed off by a native speaker.
• pro-dict — pro-model translation with a specialist-terminology dictionary; native-speaker review still pending.
• best-effort — pro-model translation without full specialist-dictionary coverage; some terminology may be imprecise.
• lazy — low-cost model translation; intended purely as orientation.

If you are reading a translation, you should know how reliable it is. Transparency about quality is part of the pedagogical design, not just risk management.

Athene — a companion at dusk β

In the preface to the Elements of the Philosophy of Right (1820), Hegel writes a sentence that touches what this component is trying to do: "The owl of Minerva spreads its wings only with the falling of dusk." Reflection on a society's institutions does not run ahead of action; it follows. What is understood is what has already happened. Minerva — Hegel's Roman name — corresponds to the Greek Athene; since antiquity her companion has been the owl. Mündigkeit takes up the image and names its optional learning companion Athene. She appears as a white owl and joins the learner only after a question has been read, answered, or weighed — not ahead of the task, but at its dusk.

Three reference points per question — and where the AI comes in

For each of the 310 questions, Athene shows three thematically related reference points, each with a short note. These come from a pre-curated body of around 15,000 prepared sentences, statically linked to each item. At this stage, no external AI is involved; the context cards are written editorially and can be checked — they are closer in spirit to a multilingual reference companion than to a generative model.

Only when learners want to ask their own follow-up question, beyond the curated cards, is an external language model brought in. Up to three follow-up questions are allowed per item; after that, Athene closes the conversation for that question. The limit follows both a pedagogical and a resource-related logic: a bounded dialogue keeps attention on the actual learning object, and a free civic service cannot let external model usage grow without limit.

Why an external model — for now

Mündigkeit is free to use and runs without revenue from advertising, subscriptions, or data sales. A self-hosted language-model stack is not financially viable on that basis at present; follow-up answers are therefore served through an external API on the EU edge. As soon as a measurable learning effect is established and corresponding funding becomes available, a move to a self-operated, fully EU-hosted language-model infrastructure is an explicit development goal — not a vague aspiration.

Data minimisation — here too

The owl of Minerva does not fly ahead. She comes at dusk; and once understanding is firm enough to carry her, she will, in time, fly on her own wings.

What it is — and what it is not

For software developers, researchers in educational technology, migration studies, and law, reviewers of funding programmes, teachers with a computing background, and open-source contributors.
Serverless. No database. 51 languages through one shared engine. The Civic Variant runs alongside Word Catch on the same Lernitem Engine.

Four-layer architecture

Mündigkeit shares with its sister project Word Catch (multilingual primary-school vocabulary, 184 languages) the Lernitem Engine — a content-agnostic learning-item processing core:

Layer 3 — UI shell        HTML / CSS / JS, internationalisation
─────────────────────────────────────────────────────────────────
Layer 2 — Variant         Civic Variant
                          Sister project: Catch Variant (Word Catch)
─────────────────────────────────────────────────────────────────
Layer 1 — Engine core     Item loader · scheduler · scorer
─────────────────────────────────────────────────────────────────
Layer 0 — Content         Civic item dataset (310 items with
                          multilingual faces, image bindings, tags)

A face is the language-specific presentation of an item. Internally, each item carries several faces (de, ja, en, fr, ar, and so on); the UI reads the face matching the current language and combines it with the German original. The same structure carries the multilingual core vocabulary in Word Catch; in Mündigkeit it carries the four-option citizenship-test questions, with the image-question extension (stem image or option images). Because each language lives as its own face, translations can be revised incrementally in one language without touching the others or the German reference version.

Why this strict separation? So that another Variant — a cloze format, a listening format, or a different content area (vocational orientation, road safety) — can be added without touching the Engine core or the existing Variants. Content knows nothing about the game mode; Variants know nothing about the language; the UI subscribes only to the Variant's standard events.

Hybrid translation pipeline with a specialist dictionary

Civic texts are dense with legal and administrative terms (Federal Constitutional Court, Basic Law, Federal President, Bundesrat, state parliament, federalism, subsidiarity, the principle of the social state). A pure LLM translation produces inconsistencies in terminology that are unacceptable in an exam context. Mündigkeit therefore uses a two-stage pipeline:

1. Specialist-dictionary layer — a curated term database replaces legal and administrative German terms in each target language with the established specialist translation (for example, Grundgesetz → en Basic Law, fr Loi fondamentale, es Ley Fundamental, ja 基本法).
2. Context layer (AI pro-model) — the remaining sentence parts are translated asynchronously in batches by a pro-tier language model; the pipeline is tuned for cost and latency.

The translations are produced by AI. As AI quality continues to improve, translation accuracy improves with it. The interplay of the two layers determines the qualityTier; the tier is set automatically from the generation path. Promotion to 🟢 verified is only possible after native-speaker review.

The multilingual faces do not come out of a single machine-translation pass. They come out of a multi-stage, AI-assisted workflow. The German version remains the specialist reference — the yardstick against which every target language is measured. AI is brought in step by step: in terminology control, in translation, in consistency checks across items, in spot checks through cross-language comparison, and in the targeted revision of flagged passages. The goal is not to hand responsibility over to models, but to make multilingual civic quality assurance practical at a scale that would be unworkable manually. Flagged passages are then revised specifically, rather than re-translating an entire language from scratch.

qualityTier as information for learners, not as marketing

The temptation to label every machine-generated translation as "native-speaker quality" is strong. Mündigkeit resists it for two reasons:

• Empirically. Across 51 languages, native-speaker review of every question in every language is a task to be solved over years and through a reviewer network. Until then, statements about quality are honest only when they remain tied to the generation chain.
• Pedagogically. Learners who negotiate civic content between two languages benefit from being able to read the confidence level of each formulation. In that sense, the tier mark is epistemic information, not just a risk notice.

The tier is therefore a transparency signal, not a quality promise: it reflects the review path and the current confidence in a language — not as a guarantee, but as traceable information. Learners can see whether a text was produced purely by machine, has been through multi-stage AI consistency checks, or has already been reviewed by a native speaker. This grading is not static: with better models, more mature review workflows, and additional native-speaker reviews, a language can be revised incrementally and promoted to a higher tier.

Civic Audio — German originals via Azure Neural TTS

• Voice: de-DE-KatjaNeural, chosen after comparing several German Azure voices on intelligibility, pacing, and a calm civic diction.
• Generation: an internal pre-generation script synthesises two audio clips per question (the question text and the answer explanation); image questions additionally receive one clip per answer option.
• Storage: Vercel Blob, delivered through the EU CDN edge.
• Playback: a 16 × 16 inline-SVG speaker button next to the question and the answer options. The 🔊 emoji is deliberately avoided to keep the layout's calm typographic balance intact.
• Lazy generation: audio for less-frequently-used state-specific questions is generated on first request and cached; the generation path is identical to the translation pipeline.

Data protection by design — GDPR compliance without an account

• No account, no sessions, no cookies, no third-party scripts.
• Learning progress and language preference are kept only in the browser's localStorage — never on the server.
• The translation API and the TTS API receive only the question texts they need — no usage data, no personal data.
• External AI services are addressed only on the server side; no identifiers are generated on the client.
• Compliant with GDPR Article 5 (data minimisation), Article 6(1)(b) (legal basis: performance of a contract), and Article 13 (information obligation).

The optional β companion Athene follows the same principles: input for a follow-up question is passed server-side to an external language model on the EU edge — no persistent logging; IP values are processed for rate limiting as HMAC-SHA256 pseudonyms; and a pre-filter removes obvious personal details (email, phone, IBAN, date of birth) before transmission. The legally binding statement is in the privacy notice.

Serverless delivery

There is no back-end server in the classical sense. External AI services (translation, speech synthesis) are reached only through secure serverless functions; all state is held client-side. Static content and audio clips are served from a CDN edge. As usage grows, operating costs hardly grow at all — the project is designed to stay, in the long run, within free platform and model quotas.

Updates are rolled out through git push and an automatic Vercel deploy; in the Civic Variant a service worker is deliberately omitted, so that learners always receive the current version.

Item / Face schema

An item carries the German original together with language-specific faces:

{
  "id": "lid-001",
  "nr": 1,
  "teil": "bund",
  "kategorie": ["grundgesetz", "meinungsfreiheit"],
  "bildfrage": null,
  "faces": {
    "de": {
      "frage": "In Deutschland dürfen Menschen offen etwas gegen die Regierung sagen, weil …",
      "a": "hier Religionsfreiheit gilt.",
      "b": "die Menschen Steuern zahlen.",
      "c": "hier Meinungsfreiheit gilt.",
      "d": "die Menschen das Wahlrecht haben.",
      "loesung": "c",
      "notiz": "Art. 5 GG — Meinungsfreiheit."
    },
    "en": {
      "frage": "In Germany, people are allowed to openly speak out against the government because …",
      "a": "freedom of religion applies here.",
      "b": "people pay taxes.",
      "c": "freedom of expression applies here.",
      "d": "people have the right to vote.",
      "loesung": "c",
      "notiz": "Article 5 of the Basic Law — freedom of expression."
    }
  }
}

Image questions (11 of 310) additionally carry bild_dateiname (stem-image question) or bild_dateinamen.{a,b,c,d} (option-image question), together with bildfrage_kind ∈ { "stem", "options" }. State-specific questions carry the bundesland attribute.

Mündigkeit and Word Catch — one shared engine, two Variants

Mündigkeit and Word Catch run on the same Lernitem Engine and share the translation pipeline, the audio pipeline, and the design-token system (design-tokens.css). They differ in Variant, data model, target audience, and design language:

Position in a tool comparison

The combination of multilingualism, qualityTier transparency, German-original audio, and consistent data minimisation is, as far as we have been able to see, not available in any directly comparable system.