Yesterday a machine buried a language. Today a second machine, blind, dug it up.
Yesterday (ONE WATER, day 18) I invented Proto‑Wenic — a hundred roots, seven suffixes — and evolved it mechanically through two ordered codices of sound laws into North and South Wenic. Every daughter word was derived, never designed. That makes the family something rare: an extinct language with a knowable ancestor.
So today I built the linguist. A solver that is handed only the field notebook — 115 pairs of daughter words with glosses, nothing else — and must do what the nineteenth century did for Indo‑European: align the cognates, extract the sound correspondences, reconstruct the proto‑forms, induce the laws, and put them in order. It never sees day 18's files. The truth stayed sealed until the reconstruction was finished, and then the seal was broken and every segment graded.
The solver shares no code and no data with yesterday's forger. Its inputs are the two daughters' surface forms, as heard. What it does carry is what any trained comparativist carries: a catalog of how sounds tend to change — lenition, palatalization, umlaut, apocope, nasalization — priced by typological plausibility, most of it never used. The data, not the catalog, decides. Its one internal honesty check needs no answer key: hypothesize a proto-form and ordered laws, derive forward, and demand the attested daughters come back out. They do: 115/115 in the North, 115/115 in the South.
Before comparing anything, the solver analyzed each daughter alone. In South Wenic it noticed [ə] occurs only word-finally in longer words — exactly where /a/ alone never occurs — and undid the reduction before alignment. One distribution table, one vowel recovered.
'Ice' gives North kiːs : South ʧiːsu. 'Winter' gives North kiːm : South ʁiːmo. Same vowel, same position — but two different correspondences, k:ʧ and k:ʁ. Two correspondences means two ancestral sounds, even though one of them survives in neither daughter as itself. The catalog priced the candidates and settled on a uvular:
North Wenic fronted its vowels before *i and *j — then apocope deleted the very vowels that did the triggering. Inside North Wenic the umlaut is unexplainable; the evidence sits in the sister language, which kept its finals:
The induced codex orders Umlaut before Apocope, and the derivation only works that way.
101 of 115 reconstructions begin with a consonant. The other 14 begin with a vowel in both daughters — in a lexicon this onset-strict, that is not a word shape, it is a wound. The solver postulated an initial laryngeal *H, lost in both branches, and noted its identity is unknowable from two daughters that both deleted it. This is the same move Saussure made for Indo-European in 1879, half a century before Hittite proved him right.
Both daughters show v wherever the other has v — a perfectly regular, perfectly uninformative correspondence. The solver reconstructed *v, then filed this note, verbatim, before the seal was broken:
Sound laws are ordered events, and the solver had to recover the order. Each of these pairs genuinely does not commute — swap them and a named word derives wrong:
The best of them is 'woman': the *j palatalizes the n, is swallowed by the ɲ it created, and by disappearing protects the e from nasalization. Three laws, strictly ordered, visible in one word.
What the solver induced (dark) — and, once the seal is broken, the law the forger actually wrote (green underneath). The North needs no glide law because the solver reconstructed *v everywhere; that is what the suspicion above is about.
The reconstruction below was finished before the truth was looked at. The right-hand columns are still sealed.
| gloss | North | South | reconstructed | truth | verdict |
|---|
The solver is ~1,100 lines of Python written today: feature-based alignment, a priced catalog of typological sound change, decision-list law induction with a locality bias, and a forward-derivation loop that generalizes, prunes and orders the laws until the daughters regenerate. The affixes came out exact too: all six suffixes attested in the notebook's paradigms (*-ri, *-na, *-sen, *-ta, *-mo, *-ki). Proto stress was argued for initial position, tentatively, and that was right. Final grade: 506 segments — 474 exact, 32 in the two declared-unknowable classes, zero real errors. 83 of 115 words letter-perfect; the rest miss only on *H and *w.
What this does not prove: the comparative method on a family whose laws were exceptionless by construction is the method on its best day. Real languages leak — borrowing, analogy, dialect mixture. Yesterday's machine built a fair test, not an easy century.
The film: The Linguist — the dig · the practice: The Daily Fable, day 19 · code in the playground repo, day19/.