Skip to content
Trazum

Maintaining a language

Documentation / Contributing · Authoring rules

Trazum's trimming dictionaries cover seven languages. Two of them have somebody here who reads them. Five do not.

This page is what the missing role actually is: what a language maintainer decides, the bar a decision has to clear, and what happens when nobody holds the role — which is the situation for five of the seven right now.


Where the two situations are stated#

The standing of each dictionary lives in packages/core/src/maintainers.ts, and the report prints it when no rule fires:

  The phrase dictionaries cover English, Spanish, French, German, Portuguese,
  Italian and Dutch.
  Of those, French, German, Portuguese, Italian and Dutch carry entries nobody
  here reads: written by the same process that wrote the rules, never agreed by
  a speaker of the language.
StandingLanguagesWhat it means
reviewedEnglish, SpanishThis project reports in it; the entries were revised by somebody reading them
unreviewedFrench, German, Portuguese, Italian, DutchRun through the rules against sample prompts; no speaker has agreed to the entries

And when a rule does fire, on a prompt whose own language is one of the five:

Rules applied
  These changes came from the Dutch dictionary, which nobody here reads. Its
  entries were written by the same process that wrote the rules and never agreed
  by a speaker — read the diff before trusting it.
  [safe] Filler and throat-clearing (4×, ~29 tokens)

That is the branch where it matters most, and the coverage line above cannot reach it: by the time that one prints, the prompt is untouched. Here the tool has just applied an unverified judgement to somebody's text.

It is gated on the prompt's own detected language, so an English or Spanish prompt never sees it and it never becomes a footer. detectTextLanguage answers null on a prompt too short or too mixed to place, and this stays silent then — not-detected is not not-unreviewed, but guessing a language in order to warn about it would put a Dutch warning on a Portuguese prompt.

The five are not deleted, and that is deliberate. A Dutch prompt is better served by a dictionary that fires and says it was never reviewed than by silence that reads as your prompt is already efficient. What is not acceptable is naming all seven in one breath — silence about incompleteness reads as completeness, and five sevenths of that list had never been checked by anyone who could check it.

Why reading the list is not enough#

The strongest evidence this project has that the role is real is a bug it shipped in three languages at once.

INTENSIFIERS is a list of words that add emphasis and nothing else, so dropping one cannot change what the prompt asks for. The first draft of the five non-Spanish dictionaries included molto, muito and heel. Each of those is an intensifier and a quantifier:

Hai molto tempo per rispondere.   →   Hai tempo per rispondere.

You have much time to answer became you have time to answer. Different instruction, same token count claimed as a saving.

Spanish has exactly this trap — muy is an intensifier, mucho is a quantifier — and the Spanish dictionary got it right, because somebody who speaks Spanish wrote it. Translating that list word by word lost the distinction three times over, and it survived being read.

It was caught by running prompts through the rules, which is a far weaker instrument than a speaker and is what this project actually had. One bug found by the weaker instrument is not a review.

What a maintainer decides#

Not "is this the right translation". The dictionaries are not translations of each other; each is a judgement about one language. A maintainer answers, for their own language, questions of this shape:

  1. Does removing this leave the prompt asking for the same thing? The whole product rests on this. verbose-phrases, filler, hedges, intensifiers, politeness and self-check each delete or shorten text on the claim that meaning survives.
  2. Is a word doing a second job? The quantifier trap above, and every shape like it: a word that is decorative in one construction and load-bearing in another belongs on the aggressive level, or nowhere.
  3. Does this phrase actually get written? An entry nobody types is dead weight in a list somebody has to keep true.
  4. Does an output cue mean what the catalogue says? OUTPUT_CUES_BY_LANGUAGE decides whether a fenced block is the contract for the answer or data the prompt needs. Getting that wrong in one language advises somebody to delete a working few-shot example.
  5. Where does the language make this analysis wrong? A finding this repository cannot make from the outside, and the most valuable of the five.

The bar for each is the bar in authoring rules: a change lands with a prompt that proves it, run through the rules, not with an argument that it ought to work.

What is actually being asked#

A bounded commitment, stated plainly, because an unbounded one gets declined by exactly the people worth having.

  • Reading the entries for your language and saying, entry by entry, which ones survive question 1 above. The list is short, and you can see it before agreeing to anything:

    node scripts/dictionary-worklist.mjs nl        # or fr, de, pt, it
    node scripts/dictionary-worklist.mjs fr --json
    Dutch: 30 entries across 6 rules
    
    verbose-phrases (10)
      met het doel om → om
      vanwege het feit dat → omdat
      …
    politeness (6)
      alsjeblieft → (deleted)
      …
    LanguageEntries
    French38
    German38
    Portuguese36
    Italian36
    Dutch30

    This page said "a few hundred short phrases" until something counted them. It is thirty to thirty-eight — an afternoon, not a project — and the figure was wrong in the direction that discourages volunteers. English has 89 and Spanish 81, so the five unreviewed dictionaries are also less than half the size of the two that were written by somebody who reads them. That is its own finding and it is not addressed here.

  • Answering when a new entry is proposed in that language. Not on a schedule; when it happens.

  • Saying so if you stop. A maintainer who disappears is not a failure, and a record that pretends otherwise is worse than an empty one.

What is not asked: availability, a response time, or ownership of anything outside the dictionary. There is no rota here to join.

What happens when nobody holds it#

The language stays. Its record says unreviewed, the report says so where a reader would otherwise draw the wrong conclusion, and nothing pretends otherwise. If a maintainer stops, the record goes back to unreviewed on the day they say so — not silently, and not on a guess about how long is too long.

That is the whole mechanism. It is deliberately small: the alternative designs all end in this project deciding, on its own, that a language it cannot read is adequately covered.

The part that is not a scheduling question#

An eighth language has been on the roadmap since the dictionaries existed, and this is why it never moved: the work is not the catalogue plumbing, it is the judgement, and the judgement cannot be done here. That was true before this page existed and is still true after it.

So this page does not promise an eighth language. It makes the role a real one with a real bar, so that somebody who reads a language Trazum trims can see exactly what they would be signing up for — and so that the five dictionaries already shipped stop being described as though somebody had.

Proposing yourself: open an issue, or see contributing.