What is known, what is not
Short version. We can count. We cannot read.
Linear A is not a literature. It is lists: a word, a number, a word, a number, a total. The part that is numbers is solved. The part that is words is not, and this page says how far each side got.
What holds up
The numerals
A vertical stroke is one. A horizontal stroke or a dot is ten. A circle is a hundred. A circle with strokes is a thousand. Write them biggest first, add them up, done. 352 of 1,880 documents carry at least one number.
Take the last line of tablet HT 13. It reads: a circle, three bars, then a small fraction sign. One hundred, thirty, and a half. The six entries above it are 5, 56, 27, 18, 19 and 5. Add them and you get 130. The scribe did too.
One quantity per line, almost without exception: 1,285 of 1,289 lines with a numeral carry exactly one, in the sign-by-sign text.
The totals
The word read ku-ro sits at the bottom of lists. It appears on 35 of 1,880 documents. Where the entries above it survive, the number after it is their sum.
We checked. In the sign-by-sign text of the corpus, 35 ku-ro totals can be tested. 10 of 35 testable ku-ro totals in the sign-by-sign text balance exactly. The other 25 fail for a reason you can point at: 15 have text missing from the span, 4 have more entries after the total, 3 have no entries left to add, 2 balance in the integers and miss by the fractions, 1 is off by one or two with no fraction in sight. Unexplained failures: 0 of 35. Each document page shows its own check.
The control is worth knowing. In Linear B, which can be read, the same test on the word for "total" passes on only 4 of 49 cases. Clay breaks. Our 10 of 35 is what a real ledger looks like after 3,500 years.
The libation formula
A set of stone vessels carries the same run of words, with small changes, as if a fixed phrase were being written out. The core phrase has sign *301 in the middle of its first word. It appears on 16 of 1,880 documents at 7 of 55 sites: 5 at Iouktas, 5 at Syme, 2 at Palaikastro, and 1 each at Apodoulou, Kophinas, Psykhro and Troullos. Of the 16 objects, 14 are stone vessels and 2 are libation tables.
What the words mean is not known. Whether any slot in the phrase names the place is also not known: we tested it and no slot does (smallest p 0.12, n = 16).
The fractions
The small marks after a number are fractions. GORILA lists 17 fraction signs with 480 attestations in total. Corazza and colleagues (2021) solved 12 of the 17 from arithmetic constraints alone: the commonest, J, is one half (137 attestations), then E is a quarter (105), D a sixth, B a fifth, K a tenth. Combinations add: J plus E is three quarters.
Our own check reaches one of those values without their paper. On HT 104, two J make a whole, so J is one half. The rest we take from them, with the citation.
The Linear B connection
Linear B is the script of Mycenaean Greek. Ventris read it in 1952. It took most of its signs from Linear A, so 94 of 378 sign types on this site share a shape with a sign that can be read. Scholars read the Linear A sign with the Linear B sound. That is how ku-ro gets its name.
It is an assumption, and a shaky one. A sign-specific argument for the transfer exists for about 20 of the 58 syllabic signs. Convention assumes all 58. A shape can keep its look and change its sound in 200 years, and it did in other scripts.
The scripts are related. The languages are not shown to be. The 138 word forms shared between the two corpora are what sign frequency alone predicts (p 0.955 against a frequency-matched null, 1,000 draws).
What does not
The language
Nobody knows what language the scribes spoke. "Family unknown" is the accurate phrase. "Isolate" claims more than 7,000 signs can show.
Why about 7,000 signs is too few
The published corpus is 7,574 signs (GORILA plus the 2024 supplement). Ours holds 6,688 sign tokens in the sign-by-sign text, and 2,261 of 8,949 slots are damage. Linear B, for scale, is 47,271 signs in the corpus we hold.
We measured how much text a decipherment needs. Take Linear B, scramble its 90 signs at random, and attack it with a model of the language. At 250 signs the attack recovers 11 percent of the sign values. At 1,000 signs, 60 percent. At 3,000 signs, 89 percent. The curve flattens there. Call 3,000 the unicity distance, the point where the text has enough redundancy to pin the key.
Linear A has 3,895 to 4,316 sound signs in its word list, depending on how you count. It clears 3,000 by a third, not by a mile. With a quarter of the slots damaged, it sits at the edge. One more number tells the real story. At exactly Linear A's size, the attack with the right language model recovers 91 percent of sign values. With the wrong language, 0.3 percent. The corpus is big enough. The missing input is a sample of the language, and no excavation supplies one.
The main proposals, and where each stopped
Greek. Read as an early Doric dialect. Each reading is offered as a clue, none has an independent check, and the words carry affixes at about five times the Linear B rate. Not shown, which is not the same as ruled out.
Semitic. Gordon (1957 onward) found about fifty lookalike words, first Akkadian, then West Semitic. Vocabulary, never grammar, and the matches point at three branches at once. A Semitic loan for ku-ro is still on the table.
Luwian. Palmer's Anatolian reading is grouped with Gordon's by the field's own surveys as formal coincidence, and nobody has taken it further.
Etruscan-related. Facchetti proposed it and ranked it himself as a suggestion that cannot be verified or refused on this much material.
Hurrian. Van Soesbergen reads ku-ro as Hurrian "again". No mainstream follower since 2015, and the same morpheme matches support an Indo-European reading equally well.
Isolate. The transliterators (Duhoux, Younger, Salgarella, Davis) offer no translation, and that is the honest position, not a finding.
Every proposal fails the same way. With 60 to 90 open syllables, words of 2 to 4 signs, and spelling rules that let you add or drop consonants, lookalike hits are what chance produces. We tested that too: meaningless strings shaped like Linear A words get an English gloss 93 percent of the time from a made-up dictionary. This is why the site glosses nothing beyond ku-ro, the numerals and the commodity signs.
What this project has killed
Our own ideas, with the number that did it.
Scribal hands from 3D scans. We tried to tell scribes apart from stroke geometry on 3D-scanned tablets (17 faces, 928 strokes). Hand attribution scored AUC 0.452, chance is 0.5, and the kill line was set at 0.70 before we looked. Dropped. The same features identify the tablet (AUC 0.882), which is a nice result for a different question.
The listing search. Eight sign types appear in the traced drawings but on 0 of the 802 SigLA document pages. We thought that proved a renderer bug. It proves an omission keyed on sign type, and nothing about the cause. Half a kill.
Sign *301 as a variant of na. A 2026 proposal that *301 is an allograph of the sign read na fails a permutation test, p below 0.0001, on the 22 word-internal *301 tokens in the word list. Its neighbours are not na's neighbours.
The sign *301 is this project's emblem because it is the loudest unknown. It appears on 272 of 1,880 documents, and on 6 of 272 documents with *301 it is stamped alone or nearly so on a small clay sealing. Every value claimed for it has failed a check, including one of ours.
Sources
- Corazza, M., Ferrara, S., Montecchi, B., Tamburini, F., and Valerio, M. (2021). The mathematical values of fraction signs in the Linear A script.
- Del Freo, M., and Zurbach, J. (2024). Recueil des inscriptions en lineaire A. Supplement 1. Etudes Cretoises 21.6. Ecole francaise d'Athenes.
- Godart, L., and Olivier, J.-P. (1976 to 1985). Recueil des inscriptions en lineaire A (GORILA), vols. 1 to 5. Etudes Cretoises 21. Ecole francaise d'Athenes.
- Salgarella, E. (2020). Aegean Linear Script(s): Rethinking the Relationship between Linear A and Linear B. Cambridge University Press.
- Salgarella, E., and Castellan, S. SigLA: The Signs of Linear A, a palaeographical database. sigla.phis.me.
- Younger, J. G. (2024). Linear A Texts and Inscriptions in phonetic transcription and Commentary, and Linear A Lexicon.
The counts on this page come from the project's corpus files and its own test scripts. The unicity test, the ku-ro audit, the formula test and the 3D pilot are each written up with their data. Write to the address on the about page for a copy.