As of 22 September 2026
What comes next
Heidi is twelve days old. So there are no quarters and no dates here — a date is a promise, and it is too early to make one. What is here is the order and the reason for it. Anything blocked says what on.
Next
Decided, and the material for it is already in the repository.
A Zurich lexicon, against invented facts
The four guards between the model and the learner catch invented RULES and empty loops — but not one invented fact.
- Asked about «Im Kauz» — a Zurich bar — the model decided Kauz was a typo for Huus and supplied the sound law k → h to go with it. Confident, plausible, invented.
- The rule is stripped today, because the pack does not list it. The wrong gloss still stands.
- This is the next piece of linguistic work and it needs a checkable word list, not a second model.
Per-area contrast data
The dialect pages tell; they cannot yet ask.
- The container shipped with the Alemannic branches. What it needs is rows somebody can cite.
- Inventing them is the one thing this product must not do — §2: this is precisely what the learner cannot check.
A second dialect
Bernese is first on the list, and the architecture is already waiting for it.
- A variety is data: one pack, one tag, one gate. No module holds a Swiss fact outside lib/variety.
- The proof is that a Ukrainian pack sits in the repository and nothing about it is Swiss.
- Zurich has to be right first. A second half-checked dialect is worse than one checked one.
Blocked — and on exactly what
Not on effort. On one specific thing somebody else would have to release.
The listening lab
Blocked on exactly one commercially-cleared dialect voice.
- Multi-talker training is the best-evidenced method available to us (Lively 1993; Clopper & Pisoni 2004).
- It cannot be built out of a Standard German voice reading dialect spelling — §7.2 and §8 both forbid that.
- Almost every Zurich corpus is CC BY-NC or has a contested licence. One cleared voice unblocks this; nothing else does.
Native review of the situations
A hundred and forty lines nobody from Zurich has read.
- Every line passed the deterministic gate, which is a claim about forms and not about how it lands.
- The pages say so themselves today. That is the honest stopgap, not the fix.
- These are sentences somebody says to a frightened person at half past six in the morning.
Deliberately not built
Standing decisions from §8, not a backlog. They are not coming later.
Day streaks and loss framing
Not «no numbers» — no number that measures consumption, and no sentence that threatens a loss.
- Sharpened on 22 September 2026 after a fair challenge: why are we so hateful towards streaks, percentages, gamification?
- Two independent questions decide it. WHAT is counted — consumption (days opened) or capability (forms you now get right)? HOW is it framed — gain («here is what you can do») or loss («don’t break your streak»)?
- What is refused is the dishonest quadrant: consumption counted, loss framed. Not measurement, and not motivation.
- What was built instead: «What you can do now», in your own space. It counts things asked at least four times and almost always answered — and it can never go down in front of the reader.
A score for pronunciation
Nobody can measure that honestly today, and printing a number would be a claim about a person.
- What is measured is how long you spoke and where the pauses were. Those are facts about the recording.
- The speaking loop ends in a room with other people. A score is what a product offers instead of that.
Promising dialect transcription
What is sold today as «Swiss German recognition» returns Standard German.
- That throws away exactly the information a learner needs.
- So Heidi transcribes what you want to SAY, in a language you already have. Not the dialect.