On September 2nd, someone signing as Recep Cinet started a language called Amadunia (ama, love, plus dunia, world). It runs on a single principle: take the easiest feature from every language you can find and leave the rest outside — "if a sound or rule is hard for anyone on Earth, it's out." Twenty letters, one sound each, always. Three hundred roots. Twenty-one settled grammar rules, twenty-six lessons, thirty texts. On September 4th he posted it to Hacker News, describing it under the title as a conlang site where every number is derived from the source repository. It got one point and no comments. The repo stands today at 262 commits, zero stars, zero forks. The language is seven days old.

That principle is a cut, and a cut always leaves a remainder: the words the rule killed. Almost every language project treats those as engineering waste and hands you only the surviving lexicon. Amadunia does the opposite — it keeps the dead words, and it books them with the price attached. Hijau, anjing, mvua are out for using letters outside the twenty, and what was lost is green, dog, rain. Fenestra, Fenster, skribi are out for three consonants in a row: window, and to write. Beli is out for sitting one sound from beri, to give, and what was lost is to buy. Luna is out for standing too close to luma, light: the moon. Matar is out because it means "to kill" in Spanish, and what was lost is to marry. The table in the README does not say "beli rejected." It says beli rejected — lost: to buy. The remainder here carries a price tag. The hardest case: ruma, house, had already been used eighty-four times in the material and was executed anyway, replaced by dom, because it stood against luma across nothing but l and r — a contrast the author ranks above how widely a word travels. Every rule in this language sits on top of a word it killed.

The remainder is load-bearing in three separate places. First, every grammar page ends with what that rule rejected. Second, all thirty texts end with what the language could not say. The fifth is a poem, and its column lists three things: stress was undecided when it was written, so the poem has no rhythm at all — read the same line two ways and both are correct Amadunia, and two readers cannot agree on the metre of one line; there is no word for "change," and the poem is about day turning into night and back, so the language cannot name the thing it is describing; and whether a line may break where it likes is still ungranted. Third, forty open questions are gathered on one page and ordered by the pressure the writing actually put on them — a word for "then" leads with six pages demanding it, while two questions have never been reached for at all. Four of the forty were not left open on purpose; they were forms the material already relied on that no rule had ever granted, and the imperative topped that list with three pages that tried and stopped, until it was granted on September 3rd. Which means the direction this language grows in is set by its own remainder rather than by its designer's plan. That is cultivation, not severing: the next round is not chiselled out of the last one but grown from its body. And when stress was finally settled on the 3rd, the poem written during the ruleless period became scannable for the first time — its closing line, Kita ama dunia una, turns out to be the only regular line in the whole poem, four clean trochees with nothing left over. The author's own note: that was luck, not craft, and it could not have been craft, because when the line was written no rule existed that could tell anyone where its beats were.

There is a rarer layer under that. The script that checks the language also generates the list of its own refusals — a page titled "What check.py refuses," 314 guarantees in 143 groups, every line the literal message the checker prints rather than a paraphrase of one, regenerated from the code, and the checker fails if the list has drifted from it. But the page also writes down, plainly, the one guarantee it cannot check: the checker cannot verify how many lines long it is, because the test harness does its mutation testing by rewriting the checker, which makes it three lines longer. In its own words, a number that describes a file cannot be verified from inside that file when the tool verifying it edits the file. That sentence is left standing on the page rather than quietly routed around. The same page records that the line count read "eleven hundred" until September 6th, when it was found to be long stale — on the page whose entire reason for existing is that the Python is too long for anyone to read.

To be clear about what is not new: building a global auxiliary language is not new at all. Esperanto has been here since 1887, with its congresses, its Wikipedia, its century and a half of sediment — that is a construct, and its beauty is the sediment a closed chisel-construct cycle left behind. Pandunia and Globasa have been at the same brief for years. Amadunia's goal is old. What is new is the object it publishes: a language whose principal output turns out to be the ledger of what it refused. There is no name for that yet — it is not lexicography, not a grammar, not a design document. And the ledger books its own misses too. The etymology audit puts European roots at 34% of the dictionary against Esperanto's roughly 95%, and by reach the largest family is Austronesian at 38.3%. Then come these lines: Sino-Tibetan, 4.0%. Japonic, 2.7%. Koreanic, 0.0%. A language whose stated promise is that every continent should hear something familiar has printed its own thinnest spot at full size. Better still is the sentence after it. The author had explained the thin spot as structural — Chinese words are short and the short space is full — and then overturned his own explanation: modern Chinese and Japanese words are mostly two syllables, which is exactly what the four-letter rule wants, and seven of ten ordinary words tested passed every rule. The borrowing reached for the classical monosyllables and stopped there. Printing the target you missed, at full size, is harder than hitting it.

Which is why seeing it now matters more than seeing it later. The moment a constructed language acquires speakers, the record of refusals gets flattened into a style guide: ruma disappears and the dictionary simply says dom; "what the language could not say" shrinks to an FAQ; the forty open questions get answered and taken down. What is left then is a construct — possibly a good one, possibly a beautiful one, but its beauty will be sediment rather than something still alive. Right now it is not that. Seven days, 262 commits, one point, zero stars. Nobody has arrived. The last line of the README invites you to follow the commits and watch a language grow in real time. That invitation has an expiry date.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics