What the language can already say.

Original writing, not translation. Every text uses only the 300 settled roots and only settled grammar — nothing is invented to make a sentence work. Where a text cannot say something, it says so at the end, and those endings are where several of the rules came from.

• Anak espera solThe child waits for the sun. · 34 roots
• Safari por paharJourney to the mountain. · 76 roots
• Lingua iniThis language. · 43 roots
• In madina baruIn a new city. · 62 roots
• UanOne. · 21 roots
• Seti dinSeven days. · 58 roots
• Surat por mamaA letter to mother. · 54 roots
• Kaifa suru caHow to make tea. · 19 roots
• Pagi in madinaA morning in the city. · 56 roots
• Mila tahunA thousand years. · 37 roots
• Anak katabA child writes. · 31 roots
• Tarik por skolaThe road to school. · 36 roots
• Kula unaEating together. · 43 roots
• Yamur aur caRain and tea. · 45 roots
• Sol laiThe sun comes. · 42 roots
• Berapa hargaHow much does it cost. · 62 roots
• Tren aur farasiThe train and the horse. · 73 roots
• Gusa in tarikAnger on the road. · 43 roots
• Kamra miMy room. · 48 roots
• In skolaAt school. · 42 roots
• Uan umurOne life. · 122 roots
• Agar yamur laiIf the rain comes. · 62 roots
• Kim paling hayaiWho is the fastest. · 50 roots
• Mama mama miMy mother's mother. · 59 roots
• Tarik por hospitalThe way to the hospital. · 45 roots
• Dom eskiThe old house. · 45 roots
• Kalima por anak-anakWords for the children. · 39 roots
• Din miMy day. · 30 roots
• Yu mau ke?What do you want? · 37 roots
• Katab haoWrite well. · 21 roots

A lesson shows what the language can do. A text finds out what it cannot.

The first story turned up four things thirteen lessons had missed — no word for and, no word for not, no again, no way to mark place. All four are now settled rules, and the story is left as it was written, from before those words existed. The poem found the one that mattered most for anything meant to be spoken: nothing in the language said where the stress falls, so two people could read the same line as two different poems.That one is settled too — the second-to-last syllable, always — and the poem is left as it was written for the same reason the story is.

Eighteen rules can be spotted in a sentence

, and they each stand in at least eight texts. That is not the same as every settled rule, and the two sentences used to be one: the claim read every settled rule is exercised here with the narrower figure beneath it, which is an overclaim followed by its own qualification. There are twenty-one rule pages. Fourteen of them are measured here, four more rows are sub-rules of those pages — existence, una, the command and the tone question — and seven rules cannot be seen by a sentence scanner at all: definiteness because it is marked by nothing, word-formation because it is an absence, phonology, pronunciation and stress because they are properties of every word rather than of any sentence, and word order because it is true of all of them. Pronouns are the exception and are simply not tracked: the scanner could count them and does not.

A question mark is two rules, and the scan counted it as one.

Asking with a question word — ke, kim, nali, kab, kaifa, berapa — is Lesson 07. Asking by tone alone has been in the lessons since Lesson 01, and the syllabus is checked for exactly that. Of the material's 259 questions, 164 use no question word, so dating every one of them to Lesson 07 charged six lessons to any text that asked one — both figures recounted here. They are two rows now, and the tone question stands in more texts than the question word does.

Possession was on the invisible list until September 6, 2026

, on the grounds that it is marked by nothing — and being marked by nothing is not the same as being invisible. A noun with a pronoun straight after it is possession and cannot be anything else, which is a shape a scanner can see; it now stands in 23 texts, second only to place. It was the rule the reading ladder needed most, because it arrives in Lesson 06 and a text was written to that lesson.

The question rule was missing from this table until September 6, 2026

, and not because it is hard to see. Its detector looked for a question mark in the sentence, and the scanner splits sentences on the question mark, so the mark was never there to find. Counted properly, the question rule stands in fourteen texts, seventh of sixteen.

• place25
• possession25
• tense20
• conjunction19
• negation19
• number18
• copula17
• tone question16
• demonstrative15
• plural13
• question12
• subordination12
• verb chain12
• command11
• existence11
• adverb10
• una9
• comparison8

And the material itself was prepared nine different ways.

Nine scans read the lessons, the texts and the phrasebook, and each prepared the text itself: five removed the whole New words section of a lesson, four kept it. The difference is one sentence and four words — a line of prose under that heading in Lesson 06 — so the repository's two headline corpus sizes were being measured over two different corpora, the sentence count over one and the word count over the other. There is one definition now, and it removes the table rows rather than the section: a gloss is not a use, and a sentence is a sentence wherever it is printed. Merged September 6, 2026, and no check enforces it — a structural rule stated precisely enough to test turned out to need four exceptions, which is a rule not worth writing.

It was missing because the test existed twice.

The reading ladder needs the same judgement — which rules a sentence uses — and had its own copy, which read the raw line and had always worked. Compared line by line on September 6, 2026, the two copies had drifted three ways: the table could not see a question, the ladder could not see a negative command, and the ladder had no number detector at all. Only the first had reached a published figure; the ladder's two gaps were latent, and the rows did not move when they were closed.

And the sentence that replaced it went stale too.

It read "Place leads at eighteen; subordination is the thinnest at five, then comparison and commands at six and the adverb rule at seven" — and against the scan that was already in check.py, place stands in nineteen texts, commands in seven and the adverb rule in eight. Only the floor was ever checked; the four named figures were prose, and three of them drifted as the texts were edited this week. The table above replaced them on September 5, 2026 and is regenerated, so it cannot.

Two rules were thin and each was given a text of its own

, which is how the floor got to five. The adverb rule stood in one text until text 11 was written for it and used it eleven times in eighteen sentences. Commands stood in two — once in text 4, and once only as an illustration of what text 8 would say — until text 12 was written for them and used eleven in twenty-seven sentences. Both rules were introduced by the lessons well after the material had started using them, and writing a page for each is what turned that round. una was the third, at 21 sentences, and text 13 answered it.

Subordination was the fourth

— and writing its text is what showed the table had been measuring the rule by half of itself. It stood at five, the thinnest of the fifteen, and text 22 was written for it on September 6, 2026. That page then pointed out that a clause can be the object of a verb with nothing marking it, so the detector — which looks for porke, kab and agar — could not see that shape at all. Taught to see it, the scan puts subordination in seven texts, and the thinnest rule is comparison at six.

1895 sentences, 449 distinct shapes

, and the fifteen commonest cover 42% of them. The leaders are RVN at 8.3% (Ta-ta lihat stela-stela), NA at 7.5% (Sol kabir) and NV at 4.8% (Anak sema). check.py recounts the two headline numbers.

1582 of the 1895 are distinct — 84%.

The corpus is not padded with repeats: the most repeated sentence appears nine times across six files and is Mi sema, "I say", which is a quotation frame rather than a sentence anybody is making. It read Mi sema at seven until September 6, 2026, and the line went stale without anything noticing, because the two headline numbers above were checked and the sentence under them was prose; it is derived now. That matters for the frequency list, which counts every occurrence: if a sixth of the material were copies of itself the curve would be measuring the copying.

Four texts headed a summary with a sentence count nobody could reproduce.

Twenty-five sentences, 45 roots opens text 14; the roots were right and the sentences were not, and the same shape stood on texts 15, 19 and 20. Measured by the convention check.py writes down — a full stop, question mark or exclamation mark inside the text's own code block — three of the four were wrong by three or four, and none of them matched the line count either. The root half was right every time, because the foot of each page states it and that figure is checked; the sentence half sat in a phrasing the sentence check did not match. A number is checked when its wording is checked, not when its subject is. Corrected September 6, 2026, and the phrasing is covered now.

The full list, and the three rules for telling a citation from an utterance that were tried and rejected, are on the briefing that owns the distinction

— and they are there rather than here for the reason the record itself makes: this page is inside the material the scanner counts, so writing the examples out would have added ten more of exactly the thing being measured. It did, in the first draft, and the checker said so.

A gap has to be a sentence that stopped.

Before a text records one, the rule page that would grant the thing has to be read. Text 10 claimed the language could not compare one duration with another; comparison.md has a section headed With verbs granting exactly that, and the claim survived a day before being withdrawn. Nothing in check.py can catch this — it verifies that a "no word for X" note is still true, because words are in a list, but a rule is not — so it is a rule for writing rather than a check.