Two places. You type them, or the backend reads them. Every five minutes it fetches the German news from tagesschau.de, takes the first several sentences of each new article, and queues them. That is where most of the lexicon comes from: current German written for adults, rather than textbook sentences about a pencil on a table.

Every sentence taken from an article carries its headline, its publisher, a link back to the piece, and where in the piece it stood. Nothing here is presented as this site's own writing. If you publish German and would rather not be read this way, say so and you will be left out.

Not blindly. A language model wrote every word of German in here. This one is better than expected: asked to conjugate helfen, it produced ihm wird geholfen for the passive, which is right precisely because helfen takes the dative and so has no personal passive.

It is also confidently wrong sometimes, and a learner is exactly the person who cannot tell. So every form, gender and plural carries two controls: right and wrong. Marking something wrong asks you for one sentence saying what is wrong with it, and that sentence is worth more than the count.

The counts show as two numbers rather than a score. One report against two approvals means something different from one report against two hundred, and a single number hides that.

Reading is free and needs no account. Analysing a sentence nobody has analysed spends money on a language model, so each account gets a daily allowance and can ask for more. Everything after that first analysis is free forever. The result is stored, and so is every word it taught the lexicon.

You can save a whole word, a whole sentence, or a single form. A single form is usually the useful one: helfen, Präsens, er/sie/es is one thing to learn, and it can be solid while ihr helft is still shaky. Forms group under the word they belong to, so a verb with twenty forms is one entry rather than twenty.

You grade each card when you review it, and the grade sets the gap before you see it next. Forgot sends it back to the start. Hard stretches the gap a little. Good multiplies it. Easy multiplies it and then some. A card you keep forgetting keeps coming back sooner, for good, rather than being forgiven after one correct answer. Re-adding a set of forms never resets progress on cards you already have.

The lexicon is shared. Every sentence, word, form and rule is readable by anyone without an account, because a shared lexicon that everyone improves is the whole idea.

Your study list and review history are yours alone. Nobody else can read them, and the storage layer enforces that rather than the application remembering to check.

Sixty-nine rules, A1 to B2, written by hand. The analysis may only choose from them. That is what stops one rule appearing under four names and the links between sentences becoming meaningless.

Where a sentence uses something the catalogue lacks, the analysis may propose a rule. The proposal is stored and linked, so nothing is lost, but it stays marked as proposed until a person approves it. You can read the proposed ones under Rules.