Opening Linear A

Linear A, the script of Bronze Age Crete (c. 1800 to 1450 BCE), has resisted decipherment for more than a century. One of the latest scholarly assessments is Brent Davis's The Undeciphered Aegean Scripts, a linguistic study of Linear A and its neighbors published by Cambridge University Press in March. Davis concludes that the structure of Minoan would tend to exclude both Indo-European and Afroasiatic "as the parent family of Minoan," leaving open the prospect of "a language with no known relatives." In contrast, my findings suggest that Minoan does have a parent family: Afroasiatic, through its Semitic branch. Today, with Plasma AI, I'm releasing the corpus, tools, and arguments so others can test that claim inscription by inscription.

The release begins with the corpus: 1,719 records from 1,603 documents, reconciled and deduplicated across GORILA, RILA-S1,¹ SigLA, and lineara.xyz. Built on those records are a lexicon of 502 entries, 439 proposed translations (with 52 more left partial where the text is damaged), and a sign list that proposes 34 values, 15 phonemic and 19 logographic.

None of it would exist without the scholars who built the record first. Louis Godart and Jean-Pierre Olivier's five-volume GORILA remains the foundation of the field, and the digital corpora that followed—SigLA, built by Ester Salgarella and Simon Castellan, and Rob Hogan's lineara.xyz—put the signs within reach of anyone with a browser. I owe them my starting point, even where my conclusions part ways with theirs.

The scripts behind the research ship with it. They reconcile transcriptions, search recurring sign groups, test each proposed value across every inscription where it appears, check the scribes' arithmetic against their own totals, and query a library of Semitic dictionaries, texts, and onomastica.

Publishing the data puts the proposal in front of its critics. Any word I read can be followed back to its inscription, its spread across the island's sites, the Semitic evidence behind it, and the rival readings it was weighed against.

Gordon's problem

For most of the twentieth century, scholars chose a language they already knew and read the inscriptions through it—Cyrus Gordon, and after him Jan Best and Fred Woudhuizen, through Northwest Semitic, Vladimir Georgiev through Greek for the tablets and Hittite for the prayers, Leonard Palmer through Luwian, and, more recently, Peter van Soesbergen through Hurrian. None won the field, but they did not leave equal evidence behind. Gordon began from the tablets' own logic: at Hagia Triada, ku-ni-su is followed by the logogram for wheat, and Akkadian kunāšu means "emmer."

I came to Gordon's problem with a new instrument: Claude Code, which I brought into my engineering work in October 2025 and soon extended with a meta-harness for testing hypotheses against the full corpus. On the night of 22 May 2026, I saw that IO Za 2, the fullest copy of the libation formula, echoes the liturgical prayers of the Hebrew Bible, and that Ventris's methods could test values for its unknown signs. Reading *301 as /na/, from Semitic √n-w-y "to dwell," proved compatible with every other occurrence of the sign, and the work moved from a hunch to a testable proposal. Through the summer I corresponded with specialists and wrote Ya Diktu, my grammar of the formula,² which responds to Davis's typological case directly, before widening the work to the whole corpus.

Linear A represented a major innovation in scribal practices of the Middle Bronze Age. Previously, scribes were required to memorize some 600 signs to record inscriptions in Akkadian and Sumerian cuneiform. The distillation of this complexity into a core set of some 90 signs (analogous to the Old Assyrian merchants' reduction of cuneiform to under seventy) made this syllabary all the more practical across an island as rugged as Crete. In my proposed reading, the syllabary primarily rendered a Central Semitic language, with the ability to also inscribe toponyms and personal names in a range of other languages. In both ritualistic and administrative contexts, Linear A records terminology that I read as Semitic, mirrored in Ugaritic, Amorite, and Biblical texts.

If so, Minoan ritual was conducted in a Semitic language, as the myth of Europa, Cadmos, and Minos had long suggested.

The Semitic family tree, the position of Minoan

Adapted from Huehnergard, An Introduction to Ugaritic (2012), p. 2, with Minoan added and Sabaic moved beside Aramaic.

Testing a decipherment

If Linear A encodes an archaic Semitic language, even frontier models know too little about it, and what they do know can mislead. Transformers are poor substitutes for databases, but they excel at writing scripts to query them. In this system, models propose searches, write the code that runs them, and sort what comes back. The evidence comes from the inscriptions and the dictionaries, through tests anyone can rerun.

A proposed value must hold wherever its sign appears, not only in the inscription that suggested it. Recurring sign groups come before meanings, and a reading has to fit the morphology and syntax around it, not merely resemble a Semitic word. Where the tablets offer independent checks—a scribe's total, an object's type, its findspot, a word's place in the formula—the reading must pass them. Evidence collection and interpretation are recorded separately, so readers can inspect the observations behind each judgment and challenge the reading without having to reconstruct the search.

Unlike an Erdős problem, Linear A has no Lean kernel to certify a result. Counts, distributions, arithmetic, and reproducibility can be checked mechanically, while segmentation, sound correspondence, and meaning stay matters of judgment, scored and open to revision. Rival languages face the same scrutiny: every Semitic match is rescored against Luwian, Egyptian, and Hurrian roots, and the margin between them is reported.

IO Za 2 and HT 31

IO Za 2 is a four-sided serpentine libation table from the peak sanctuary on Mount Iouktas, inscribed in two lines that wrap around its faces in the Neopalatial period. It carries the fullest surviving copy of a prayer that recurs at least thirteen times across seven sites:

A-TA-I-*301-WA-JA · JA-DI-KI-TU · JA-SA-SA-RA-[ME · U-NA-KA-NA]-SI · I-PI-NA-MA · SI-RU-TE · TA-NA-RA-TE-U-TI-NU · I-DA-[

The brackets mark signs lost to a break and restored from the parallel copies. U-NA-KA-NA-SI is spelled out in full at Kophinas, Palaikastro, Syme, and Troullos. Read as Semitic, the prayer can be rendered:

I make myself a dwelling—O Diqqitu!—in righteousness. So, behold my gathering in the presence of sacred song. May you be favorable toward us.

The argument rests less on any single word than on the paradigm. Across the formula's copies, the opening verb changes its first syllable: A-TA-I-*301-WA-JA at Iouktas, with ʾa- for "I," and JA-TA-I-*301-U-JA at Apodoulou, with ya- for "he." These match the Semitic imperfective prefixes, set before the -t- of a middle-reflexive stem, as in Hebrew hitnabbēʾ. Where an alabaster cup from the same sanctuary turns to the Goddess, TA-NA-I-*301-U-TI-NU, "may you answer us," the prefix becomes ta-, "you." JA-DI-KI-TU opens with the vocative yā, the same particle that introduces JA-SU-MA-TU, "O Šumāt!", on a table at Syme. The closing petition marks its subject and its object together: ta-, "you," the Goddess, and -nu, "us," the worshippers who address her.

Davis reads the same alternating syllables as long, non-Semitic affixes. The two readings make different predictions. If the prefixes are Semitic person markers, they must agree with the rest of their clauses, the "I" of the opening verb with the speaker's "my gathering," the "you" of the petition with the Goddess it addresses. The parallel copies let readers inspect the proposed person-marker analysis across recurring forms, including the segmentation and translations on which it depends.³

A prayer is one kind of evidence. An inventory is another, and it constrains a reading from outside the language. Davis notes ku-ro, "total," as a word whose meaning context alone could fix. HT 31, from Hagia Triada, goes further: its scribe drew each vessel and wrote its name onto the drawing. Ventris and Chadwick saw in 1956 that "the difference in language is obvious"—these are not Greek words. Gordon, and Robert Stieglitz after him, read them as Semitic: SU-PU on a bowl is Ugaritic sp and Hebrew sap, "basin," and I read QA-PA₃ on a cup as Hebrew qubbaʿat, "goblet." Any proposed decipherment has to survive both tests, accounting for the scribe's vessels, totals, and personnel as readily as for the Goddess and her song.

The kaphtor repository

Everything I've described lives in one repository, kaphtor, after Caphtor, the Hebrew Bible's name for Crete. The code is released under Apache-2.0 and our original data and writing under CC BY 4.0, so anyone may reuse, rerun, or contest them with attribution. Third-party material keeps its own terms.

The Linear A half begins with the corpus. It builds on lineara.xyz and SigLA, whose readings we checked against GORILA, John Younger's transcriptions, and the other primary editions, correcting many along the way. Beside it sit the lexicon, sign list, and translations, and the tools that test them: sign and formula-position search, a paradigm tester, the accounting checks, and the root-match engine that scores Semitic readings against their rivals.

The comparative half began as a reading list. I had Claude Code gather every Semitic dictionary and onomasticon I could legally acquire,⁴ run them through OCR, and build a Python engine for searching them, questioning them, and setting their entries side by side. At its center is a database of 3,799 Semitic roots, traced through their reflexes in HALOT, the Chicago Assyrian Dictionary, Lane's Arabic lexicon, DULAT, Leslau's Geʿez dictionary, and a dozen other standard works. I know of no larger Semitic roots dictionary that anyone can download.⁵ Around it, the Ancient Near East (ANE) package queries Hebrew, Aramaic, Akkadian, Ugaritic, Egyptian, and Arabian sources, builds root dossiers stamped with their provenance, and keeps primary texts apart from dictionaries, OCR, and machine classifications.

Each source is included according to its licensing terms. For protected dictionaries, we provide the cited form, source abbreviation, and page number so readers can locate the entry in the original. Over time, we plan to replace these references with entries drawn from public-domain dictionaries. Share-alike material retains its original license. Sources that cannot be redistributed are omitted or supported through tools that let readers retrieve their own copies, and scanned books are not included. The Claude Code skills used to run the research tools are included, so others can work with the same research environment I built and used.

Decipherment as collaboration

Progress on Linear A and other undeciphered scripts, such as the Indus script, will depend on researchers bringing different expertise to shared evidence. A collaborative workspace lets scholars and agents compare readings, test competing explanations, and preserve what they learn for others to build on. By releasing my Linear A corpus and research tools, I hope to make that work easier. I invite researchers, independent and academic alike, to bring their own methods, challenge my conclusions, and pursue interpretations I may have missed.

AI is expanding what researchers can investigate, from mathematics to ancient languages. An agent can search thousands of records, compare variant transcriptions, recover citations, and help test a hypothesis across a corpus. Making sense of those results draws on expertise developed over years: recognizing a damaged sign, weighing a sound correspondence, or recalling an article or conversation that changes the interpretation. I see these new tools as a way to extend that scholarship and give researchers more room to follow difficult questions. Their outputs deserve the same scrutiny as any other contribution to the field.

I welcome anyone curious about these questions to explore the material and contribute their own sources, readings, and tools. Different perspectives can help uncover what I have missed and open new lines of inquiry. The roots database and comparative philology toolkit can also support research across Semitic languages, well beyond the study of Crete.

Start at github.com/plasma-ai/kaphtor: clone it, rerun a query, and follow a word from its inscription to its proposed reading. Report a misread sign as an issue, or submit a rival analysis or addition as a pull request. I extended my tools with Plasma's open-source projects and Fractal, an internal workspace for people and AI agents. Fractal let me coordinate teams of cloud agents across parallel lines of research, delegate tasks, and build on a shared body of knowledge. Evidence, findings, and unresolved questions remained available for collaborators to inspect, challenge, and extend. Plasma is now opening Fractal to a small group of academics interested in continuing this work. Request access here. The corpus, the proposed readings, and the tools that test them are now public.

The tablets have kept their silence long enough, and the work of reading them aloud is now open to us all.