A Bronze Age script was cracked without anyone knowing what a single sign said. First the signs were sorted into a grid by pure structure — which share a vowel, which share a consonant — with no sounds at all. Then four guessed place-names pinned a handful of rows and columns, and signs that appeared in no place-name came out forced — and what they spelled was Greek. One of those forced words was tripod. Then a tablet no one could yet read turned up with that very word, beside a three-legged cauldron.
Below is a corner of the grid: a column for each vowel, a row for each consonant. Every cell holds one real sign, shown only by its catalogue number. No sound is assigned to any of them. All that is known is the structure, which signs share a row, which share a column.
Four sign-groups turn up again and again, and only on tablets from Crete. That was Ventris's gamble: they are the island's four best-known towns. You make the match. Each town you pin fixes the consonant of a row and the vowel of a column, and any sign caught in a pinned row and a pinned column comes out forced, even though it sits in no town-name. Get it right and the forced cells spell Greek. Get it wrong and the grid contradicts itself, and speaks gibberish. Only one reading of four town-names is coincidence-proof.
Not “AI art.” This is standard, textbook history of the decipherment, made operable. Nothing here is an original finding; the only new thing is the form, the grid run as a machine you drive. You place the four town-names yourself, and the propagation, the contradictions, and the Greek that falls out are all computed live in front of you. That the true match is the one and only reading of four town-names that is free of contradiction and spells Greek is checked offline, over every possible matching, in research/the-grid-that-spoke-greek/verify-bench.mjs (21/21). The interactive is a faithful subset of an ~87-sign syllabary; the propagation it shows is true to the method, smaller than the historical one.
Sources.