How Heidi is built, and how to check it

Paste something a Zurich speaker actually sent you. Heidi explains it — what was said, which words did the work, and why the sentence is shaped the way it is. Then it offers the words worth keeping, and brings them back later on a schedule built from the spacing research rather than from a marketing calendar.

Around that sit the things a learner needs next: a grammar section of eleven topics, eighty-three words chosen because they block sentences rather than because they are common, fourteen scenes of what is actually said during a working shift, and five hundred and fifty-three practice questions generated from that same checked material.

Nothing in the product invents language. Every question, every example and every gloss is assembled from material the variety pack already vouches for, which is the constraint that makes the rest of this document possible.

Swiss German is a family, not a language. The dialect of Zurich and the dialect of Valais differ enough that Swiss television subtitles one for the other. Teaching «Swiss German» in general is therefore teaching nothing in particular — so Heidi teaches one variety properly and says which one, everywhere, in the interface and in the data.

It also has no settled orthography. There is no authority to be right against, which means a product cannot mark a learner's spelling wrong without inventing a standard — and it means dialect exists mostly as speech, so the written material a model could learn from is thin and largely research-licensed.

Those two facts are why the market looks the way it does. Most voices sold as Swiss German are Swiss STANDARD German: the written standard read aloud with an accent. It is a different language from the one at the lunch table, and a learner buying it has been sold the thing they already understand. The economics of that are well described — a buyer who cannot judge quality before buying degrades the whole market — and it is the reason everything below exists.

Everything Zurich-specific in this product lives in one file: the variety pack. The forms, the rules, the grammar topics, the vocabulary, the articles, the sound correspondences, and what vouches for each of them. No module elsewhere holds a Swiss fact.

That is what makes the product reviewable by the people qualified to review it. A native speaker can read the whole of what Heidi asserts about Zurich German in an afternoon and mark what is wrong, without reading a line of code. A linguist can check a rule against an atlas. Neither has to take our word for anything.

It is also what makes a second dialect a pack rather than a rewrite. The proof that the seam is real is that a Ukrainian pack sits in the same repository and nothing about it is Swiss.

A deterministic gate sits between the model and the learner. Every generated line is checked against the pack's own list of forms belonging to other dialects — as code, with a severity threshold, not as a second model asked politely. A flagged line is marked rather than quietly removed, so anything drifting stays visible.

Three further guards sit beside it, each added after a specific failure. A sound correspondence may only be cited if the pack vouches for it. A word glossed against itself is dropped. An answer that loops is refused and retried rather than shipped. Together they remove the failure modes that matter most here: a fluent, confident explanation that is wrong in a way the learner cannot detect.

What the gate does not do is check a single fact. It governs rules and form, not whether a particular gloss is right, so a model can still be confidently wrong about one word. Closing that needs a Zurich lexicon to check against, it is the next piece of linguistic work, and it is on the roadmap rather than implied to be done.

The specification carries a list of sentences this product may not say, and it is enforced by tests rather than remembered by people. No promise to transcribe dialect. No score for pronunciation. No measurement of learning that is really a measurement of usage.

Each is refused because it cannot be made honestly today. Nothing transcribes this dialect reliably, so dictation writes what you want to SAY in a language you already have, and says so. Nobody can assess pronunciation in a variety with no standard, so the speaking page reports how long you spoke and where the pauses fell — facts about a recording rather than a verdict about a person.

Publishing that list is the point. It is the part of a product that is cheapest to write and most expensive to keep, so it is kept where it costs something: on the public roadmap, beside the plans.

Words you keep, their review schedule, your recordings and what those recordings measured all live in your browser. Not as a policy that could be revised later — as an architecture in which the server was never sent them.

Where audio does leave the device it is for one named purpose, stated on the same screen rather than in a policy page, and only for the variety where transcription is honest. In Zurich German nothing is sent at all.

For an institution the relevant limit is this: there is no data-processing agreement with a model vendor today, so Heidi is not suitable for material covered by professional confidentiality. That is a contract to be signed rather than a problem to be solved, and it is named here because it is the first question a procurement officer asks.

Read the variety pack — one file, and the whole of what we assert about Zurich German. Compare the technology page against the published word error rates; the best verifiable figure in the field is 12.1 %, and those weights are not released. Run the test suite: eight hundred and twenty-two tests, and the command is in the readme.

Then use it against something we do not control. Paste a message from a real Zurich speaker and see whether the explanation holds. That is the measurement that matters and the one we cannot stage.

If a line is wrong, tell us. The situation packs were written for this product and machine-checked for Zurich forms; the pages say plainly which of them a native speaker has reviewed. A correction from somebody who grew up with this language is worth more to us than any feature on the roadmap.