Yet another vibe coder trying to solve Linear A.

Earlier this month we audited a study that had gone viral: a reading of a small stone offering table from Crete, posted by its author himself, which turns the text carved around it into a prayer to a father-god and two deities. The author is Michael Schümann. What we found then was that every claim our data could check came out exact, and that everything giving the text its meaning sits where no check can reach. This entry is the other half of that job, and it starts from the opposite move. We granted him everything.

The prayer does not stand alone. It belongs to a set of seven papers, all posted by him on Zenodo, a public repository, without the independent expert review a journal would arrange. Together they give four old signs new sound values, build a word-by-word vocabulary, and set out about twenty pieces of grammar: endings that mark a plural, a possessive, a place, and so on. That is a large thing to claim, and it has a consequence that can be measured. An ending that really is an ending has to behave like one everywhere, not only in the nine inscriptions his papers discuss. So we assumed the whole set is true and asked what the rest of Linear A, the undeciphered writing of Bronze Age Crete, ought to look like. We worked from the locked copy of 1,721 surviving inscriptions this project always uses, fixed to one exact version so anyone can redo the counts.

Assuming someone is right is easy to do badly, so the rules came first. We tested only the rules he states for Linear A in general, twelve of them, each backed by a quotation from his own text; everything he says about one single inscription was left out. We wrote the decision rule down before running anything: how many matches would count as a real trace and how few would not. Then two fresh reviewers, brought in to attack our test rather than his theory, filed 34 objections, and we accepted all 34. Several went in his favour. One rule of his that covers all the inscriptions, and that we had missed, was added to the list of things to look for. And where two of his papers read the same sign in two incompatible ways, once as a plural ending, once as a possessive, both readings were carried through the whole calculation instead of one.

One admission belongs here too. To check that the test was feasible at all, the reviewers had to compute the random comparisons before we froze the design, so what we locked was the decision rule, not our ignorance of the numbers. We said so inside the frozen document, and at the review every result of ours has to pass before anyone interprets it, each rule we had rejected along the way was recomputed afterwards, to show in public what it would have given. The run itself was carried out by someone forbidden to interpret anything. He stopped twice on genuine ambiguities in our own frozen wording, both settled in a way that gave the same answer under every reading available; twelve counts pinned in advance all came out exactly as declared; the whole thing was executed three times and came out identical down to the byte, with 10,000 random reshuffles behind every comparison.

The first of the two results uses his own logic. On the offering stone he restores a missing ending on one word because that word without the ending turns up on its own elsewhere. Turn the argument around and it becomes a test: if his endings are real, the surviving words should come in pairs: the same word once with the ending and once without it. We took the six endings he commits to for Linear A as a whole and hunted those pairs among the 647 surviving words of more than one sign that are complete, meaning no damaged sign anywhere in them. We found 7 pairs, among them i-da-mi beside i-da, and ka-pa-qe beside ka-pa. Shuffling at random which sign ends each word produces about 3 such pairs on its own. To call the trace real we had asked for 9. Seven is what chance alone turns up about one time in sixteen, which is nowhere near enough. The weight of the test lay outside his own material, and it fell there: of the twelve pairs the run turned up in total, counting those formed by the prefix he posits and one spelling variant, exactly one comes from his nine inscriptions.

The reason for that thin harvest is not necessarily his theory. It is the state of what survives. Most Linear A words appear exactly once, so the plain form without the ending, the one his grammar needs, is usually not there to be found. Even a common word-final sign like -na, which turns up 47 times at the end of a word, yields only two usable pairs. The same thinness shows through his own wording. The ending he describes as appearing frequently throughout Linear A closes 13 words out of the 3,957 that survive. The standalone element he says occurs multiple times throughout is found 31 times, of which 25 are one contiguous batch of clay sealings from a single archive room, leaving six independent occurrences. We report that as a fact about how little material there is, not as a verdict on him, and it is why our answer is “we cannot tell” rather than “he is wrong”.

Our own mistake belongs in the middle of this, because it went his way. One of the two contested readings of that sign is the more charitable to him, and under it the count rises to 10 and just touches the bar. We would have reported that had it stood. It did not stand: the review found that our own reshuffling scheme made the comparison too easy, and with the scheme corrected, chance alone produces about 6 pairs rather than 3. Under the corrected comparison neither reading clears the bar. The error was ours, it was in his favour, and fixing it removed the one branch where his grammar left a visible mark.

The second result points the other way, and it carries the same weight as the first. The claim he repeats most across his papers, and states as a possibility rather than a certainty, is that Linear A had no sign for a sound like the sh in ship, and that scribes may have written that sound by doubling a sign: sa-sa, as in the shrine word ja-sa-sa-ra, and ti-ti, as in ti-ti-ku. Counted on those same complete words, 13 of them contain one of the two doublings, where reshuffling the signs at random gives about 3. A gap that size comes up by chance about once in ten thousand times. The control is what makes it worth reporting: every other sign in the script, taken together, doubles 15 times where chance would expect about 28. Doubling is rarer than chance everywhere except on exactly the two signs he names.

That pattern needs the same caution as the first result. Much of it rides on words that repeat: collapse every family of repeated words to a single instance and the 13 fall to 7, which chance would produce about once in 64 times, under the strict bar we had set. The single step that pushes it under is treating the place name ti-ti-ku and its longer relative i-ti-ti-ku-ni as one item rather than two. And what is real here is a fact about the script, that two particular signs get doubled far more than all the rest. It fits his explanation. It does not show that the doubling stood for the sound he says it stood for, and it says nothing about which language is being written.

The closest thing to a surprise came from three words. The reading of the offering stone turns on its long final word, ta-na-ra-te-u-ti-nu, which he cuts into pieces meaning “in the sanctuary of the god Tinu”. The same mountain shrine on Juktas left two more words built the same way, on two other stones, and his papers mention neither. So we asked whether his own grammar can take those two apart. We built an inventory of 46 entries out of everything his published papers state, and inside it no complete cut of those words exists: the best we reach covers 6 of their 7 signs, and the one left over is a single u. Then we noticed what we had left out. His paper on the offering stone floats, as a possibility, an old final -u meaning origin or belonging, and we had not put it in the inventory. Add that one entry of his and a complete cut appears. The limit we hit was our list, not his grammar, and for that reason this probe is recorded as undetermined, not as a failure of his reading.

Once again, on things that can simply be looked up, he is exact. Every word his papers say can be found outside their nine inscriptions is there, confirmed sign by sign: ti-ti-ku and i-ti-ti-ku-ni on two account tablets from Haghia Triada, wi-sa-sa-ne on a tablet from Khania, sa-sa-me and i-ku-ri-na on two more Haghia Triada tablets, di-di-ka-seon a stone jar from Zakros. Three of his editions differ from our copy over damaged signs, and one object of his set, a gold hairpin, is not in our locked copy at all. We leave those open instead of scoring them, because settling them needs the original scholarly edition in hand, and last time a difference of that kind turned out to be our own data source flattening the editors’ work, not his error.

So what does the rule we wrote in advance say? Undetermined. Not supported, not contradicted: undetermined. We should be candid that the harshest verdict on the menu, the trace is absent, was out of reach from the start, because with material this thin any negative would have been too weak to trust; the real choice was always between his grammar showing up and our being unable to tell. Every check the reviewers ran to try to move the verdict left it where it was. It says nothing about his four sign values, nothing about his proposal that an early Indo-European language lies over an older one in Linear A, nothing about what he argues for any single inscription, and it refutes none of it.

What it does say concerns the size of what he committed to. Grant the whole theory and the surviving inscriptions are still too thin to carry the print of its grammar, while the one thing it names that we could measure, the doubling, is real but fragile, and silent about sound. A reading that explains the nine texts it was built from, and commits to almost nothing measurable outside them, cannot be told apart, on this material, from a reading built on those nine. We do not mean that as an accusation. It is what the audit found on one stone in the previous entry, counted this time across everything else that survives.

Earlier this month we closed a seven-round stretch of the project and opened a bolder one, still held to the same rules: every question is written down before it runs, every result passes a mandatory review before we read anything into it, and a no is published as a no. This was the first batch of the new phase. It asked five questions about how words behave inside the surviving Linear A inscriptions, the undeciphered writing of Bronze Age Crete. We worked from a locked copy of all 1,721 of them, fixed to one exact version so anyone can redo the counts. Most of the ones long enough to study come from a single place, Haghia Triada, a site that alone left almost two-thirds of all surviving Linear A: nearly six in ten of the words we counted are from there, and nearly all of those sit on account tablets, the Bronze Age equivalent of a bookkeeper’s ledger.

The five questions were frozen together in one block before a single one of them ran, so no choice could be nudged to fit a result we had already seen. That discipline held up under strain: the person running the work stopped partway and flagged six loose ends the review round had left open, and all six were settled and locked before anything ran. The whole run was then done twice, with the randomness seeded two different ways, and came out identical down to the byte. Three fresh reviewers, none of whom had built the analysis, checked it and passed it with a set of binding cautions about how far the one positive result may be pushed. Those cautions shape every sentence below.

One of the five lit up. The question was whether the place a word takes in a list depends on which word it is, and the answer is yes. The measure behind it asks how much knowing a word tells you about where it sits, at the start, the middle or the end of its list. Here is the catch, and it is the heart of this entry. That measure climbs high on its own, just from there being hundreds of different words to place, even when you shuffle those words into random order. Shuffled at random, it already reads about 689 out of a raw total of 740. Only the last thin slice, about 50, roughly seven per cent above that floor, is real structure. So we did not report the big number. We reported the slice.

That slice is real: a result this far above the shuffled floor would turn up by luck only about one time in ten thousand, and it was the only one of the five questions to clear the strict combined bar that asking several questions at once demands. But real is not large, and it is not what a casual reading would hope. Look at what drives it and it is plainly the layout of an account list. The signs that stand for goods cluster in the middle of a list; the words that head a list or total it sit at its edges. That is ledger formatting, the same reason a modern invoice puts the total at the bottom. It is not word order in a sentence, and it is not grammar. We claim it only for these account documents from Haghia Triada, not for Linear A everywhere.

The other four questions are worth seeing together, because the shape of the batch is the real story. One asked whether set phrases repeat, the way stock legal wording does; the answer was a clean no, and the test was sharp enough that it would have caught even a modest number of them had they been there. One asked whether the order of words carries information; that came in under the bar we had set, and the test was not sharp enough to call it either way, so it stays open. One asked whether particular goods get listed together; five pairs do tend to share a tablet, but only as leads, not findings, and we will come back to why. And one was a prediction locked in advance.

The goods-listed-together result is held back for three plain reasons. It misses the strict combined bar by a hair. Its single strongest case rests on one tablet from Khania, not on a pattern spread across the corpus. And the pairs are just goods co-listed on about five account tablets from the same site, which is what you would expect of an inventory, not evidence that two words belong together in the language. So they are leads for a future dedicated test, nothing that has been found.

The prediction locked in advance is the sharpest test the batch ran, and it is the one people who follow this project ask about. Before the confirming check was run, we wrote down a top-ten list of the strongest patterns from the exploratory half and committed it; the timestamps prove the prediction came first and was never touched after. On fresh data held back for the purpose, none of the eight patterns that could be checked kept its direction. That is not a refutation and it is not a clean no: either the patterns do not carry over, or there was not enough held-back data to tell, and this test cannot separate the two. There is a twist worth admitting: the one result that did survive, the positional signal, was never in that frozen top-ten, so it was never put to this forward test at all.

The net of the whole batch is quiet. Word positions in these tablets are not random, which is the single powered, honestly corrected positive we can stand behind, but everything about it points at the layout conventions of Haghia Triada account documents. Nothing in these five questions separates “Linear A as a language” from “the format of a ledger,” and the one attempt to carry any pattern beyond that single site found nothing it could test. If you can read these tablets better than we can, the reply box is open.

The stone is known to specialists as IO Za 2. It is a small offering table, a shallow dish of dark green stone just over four centimetres across, found at a peak sanctuary on Mount Juktas in Crete and made roughly 3,500 years ago. Around its four sides runs one of the longest texts we have in Linear A, the writing of Bronze Age Crete that no one has ever deciphered. Two of its faces are damaged. This June, Michael Schümann proposed that the text is a structured prayer, and gave a full translation:

“Father of the Bull-Contest of Dikte, weep the tears; may the libation offering flow in streams in the sanctuary of the god Tinu and Ida-Mate.”

In his reading, Tinu is an ancestor of the Greek god Dionysos and Ida-Mate an early form of the goddess Demeter. The proposal was self-published, meaning the author posted it himself on Zenodo, a public repository, without the independent expert review a journal would arrange. Through August it has spread across social media and news sites, often illustrated with a computer-generated picture of the stone.

We did the one thing the excitement skips: we separated the parts of the reading that can be checked from the parts that cannot, and ran every possible check on the first group. Our checks used a locked copy of every surviving Linear A inscription, 1,721 of them, fixed to one exact version so anyone can redo the counts. For the damaged spots we reopened the plates of the inscription’s original scholarly edition, published in 1985, from the free scans held by the French School at Athens.

Start with what holds up, because a good deal does. Every look-alike word the reading leans on is real and sits where the study says. Two divine-sounding words it builds on, i-da-ma-te and da-ma-te, appear on two objects from Arkalochori and on a tablet from Kythera; a related word turns up at Palaikastro; a phrase read as “the libation offering” recurs in the same slot at three different shrines. The long final word the whole reading turns on is genuinely one of a kind in everything that survives. And the descriptions of the damage, what can still be read and what is lost, match that 1985 edition faithfully. On one point the old edition was actually more careful than our own data: our source had recorded as legible a couple of signs the 1985 editors only restore by guesswork. That is our error, not the study’s, and we log it here.

Now the other half, where the reading gets its meaning. It rests on cutting that unique final word, ta-na-ra-te-u-ti-nu, into pieces and translating them: te-u as “god”, ti-nuas the god’s name, Tinu. But the same shrine left two more words built the same way, and the study mentions neither. One shares the same beginning and the same ending with a different middle; the other keeps only the ending. The part that stays constant across all three is that ending, -u-ti-nu, and the study’s cut falls right inside it.

The reading also depends on giving old signs brand-new sound values. One sign, numbered 301 in the standard list, is read as “bull”. In the surviving inscriptions that sign appears 272 times, and in 238 of them it stands completely alone, as a word on its own in inventory and account documents; only nine of its appearances fall inside the prayer formula the reading comes from. The study does not discuss where else the sign turns up. That value, and a second reread sign the prayer needs, are argued for only in another paper by the same author, itself self-published, plus a third that has been announced but not yet written.

Then there are the meanings themselves. Each translated word is matched to a single word in one other language, chosen one at a time: Old Lithuanian for “tears”, “weep” and “let it flow”, an ancient Anatolian language called Luwian for “sanctuary” and for the offering, Mycenaean Greek for the god’s name. Only one of the nine, “god”, has a wide comparison across several languages behind it. “Tinu is Dionysos” rests on a single shared syllable. And the closing “and Ida-Mate” is almost entirely rebuilt: the original edition reads only the first two signs plus a faint trace, then a gap, and restores nothing, while the sign that would carry the word “and” is, in the study’s own words, not preserved.

So we cannot prove the reading wrong, and we do not claim to. The meanings live in exactly the place no check can reach, and that is the honest result of the exercise: everything that can be tested against the stones is solid and was already known, and everything that makes this a prayer to an ancestor of Dionysos sits in the untestable half. A viral headline collapses that distinction into one thrilling sentence. A reader who meets the claim in a feed deserves to see the seam. If you can read these signs better than we can, the reply box is open.

This project never set out to decipher Linear A. Its writing has resisted a century of attempts, and one more guess would add nothing. The aim was the opposite of a guess: to take what people claim about these inscriptions and test each claim as strictly as the thin evidence allows, so the space of things one can honestly say gets narrower and cleaner, not wider. The very first entry put that promise on the line by trying to break the project in its opening week, and something did break, a database we had counted on, which we replaced and kept as a cross-check. The rule from that day held for everything after: write the test down before running it, and publish what comes back, flattering or not.

Here is the shape of it. Twenty-four entries stand before this one. Behind them sit fifty-nine tests, each a specific question we committed to in writing before we were allowed to run it, so the recipe could never bend to the answer. And into the machine went seven separate collections of writing, our own Linear A and its deciphered cousin Linear B among them, together with a few known-language yardsticks brought in for comparison. Each collection is counted in its own natural unit, inscriptions or sentences or dictionary entries, and those units do not convert into one another, a point worth holding on to.

Most of what we tested did not survive, and that is the honest headline. A published rule that a certain word order in these inscriptions is fixed turned out to have five clear exceptions, though it does hold across the seventeen cases at its core. A search for hidden prefixes and suffixes dissolved the moment we compared it against shuffled words. A computer that measured every sign shape found nothing to change in the standard list of signs. A hunt for Linear A words that reappear as names in Linear B came back empty. Our very first positive result, an apparent match between the two scripts, lasted nine hours before our own review turned it inconclusive. And a whole apparatus we built to judge other people’s decipherment claims by a fixed mechanical rule broke five times over and had to be scaled back to a plain reasoned reading. None of these are defeats. Each is a claim that will not now be made on false confidence.

The single most important thing the project did was catch itself in an error. Early on we announced that thirty-two of forty pairings of a word with a kind of goods occurred together far more often than chance would explain. An outside check found the flaw: we had chosen those forty pairs by looking at the very numbers the test was then supposed to judge, which is not a test at all. We redid it the honest way, declaring a field of 1,728 pairs in advance, and the count came back at nine. The data had not moved a single row. What moved was our count of our own questions, corrected in public, in the entry that reports it. We turned the same suspicion on our own prose, too. A line-by-line audit of the claims written on this site examined 110 of them: 86 held up against a source, 19 could not be anchored to anything, and 5 were contradicted by our own files. The largest fix was a piece of textbook history we had backwards. We had written that Michael Ventris turned a set of sign relationships into a grid; the grid was Alice Kober’s, and the credit for it is hers.

Laid out by outcome, the whole ledger of fifty-nine tests looks like this.

What is left standing is short, and each item comes with its own brake. Some signs really do prefer the front or the back of a word, more than any reshuffle of the data would produce: one sign appears 160 times in first position and 15 times in last, and twelve signs pass the strictest bar we set. That is a fact about where a shape sits in a word, nothing more; these are sign identities, not prefixes, and it says nothing about what they sounded like or meant. A small set of frequent picture-signs turn up next to number-signs far more often than the one-in-four background rate, four of sixteen surviving strict correction, which is the fingerprint you would expect from accounts and ledgers, and it says only that the two kinds of sign keep company, never that any quantity has been read. One old claim about word endings came out consistent with its predicted direction, but that direction was fixed by the setup and published before the run, so the run tested our coverage and nothing else; it must always be cited next to its withdrawn twin, the first attempt that had too little data to speak, and its honest gap, once you compare like with like, is about two parts in a hundred, not the larger figure a raw count suggests. And a single sign shape survived the inventory test as a possible split, but on a weak secondary measure and hemmed in by circular reasoning, so it stays an open thread and was never promoted to a finding.

Step back and the tally itself carries the warning. Of the fifty-nine tests, thirty are not wins and were never meant to be: they are censuses, audits and plain counts, the groundwork a claim would need, not a bet against chance that was won. Only three turned up a signal that beats chance, and even those, as above, travel with caveats. Nothing in any of it reaches the language behind Linear A or names a relative for it. When we finally set Linear A beside four other scripts and measured them all the same way, they sorted themselves by writing system and by the kind of document, an inventory reads unlike a prayer, not by language, and Linear A landed in no special place. The profiles track the script and the genre, not the tongue.

So the arc that ran through these seven rounds is closed, and the honest summary of it is modest: a lot of careful noes, a few small yeses kept on a short leash, and a habit of correcting ourselves in the open. That habit is the one thing we are sure is worth keeping. From here the project changes gear rather than ends. The next stretch is deliberately exploratory: we will chase newer and bolder ideas about what these inscriptions might hold, reaching further than the safe questions we have asked so far. The controls do not loosen to make room for the ambition. Every new idea will be frozen in writing before it runs, will pass the same required review by people brought in cold, and will be published exactly as it comes back. We are promising the search, not the discovery. It is entirely possible that we look harder than ever and find nothing that survives our own checks. If that is how it goes, you will read that here too.

A couple of entries back the two scripts finally sat side by side in the same machine. This is one of the first questions we put to it, and it is an old one. In a 1990s article, a scholar of Aegean scripts, Gareth Owens, wrote a single line: more sign groups end in Consonant+A in Minoan Linear A than in Mycenaean Linear B. Put plainly: many signs in these scripts stand for a whole syllable, a consonant followed by a vowel, sounds like na, ra or ja. A sign group is a run of signs written together, what we treat as a word. Owens was pointing at the vowel that ends a word, and the tail of a word is a clue to how a language is built, the way English endings like -ing or -s are. If one script leans toward words ending in consonant-plus-a and another does not, that is a difference you can count.

To count it you first fix what counts, and write it down before you look, so the recipe cannot bend to the result. A word is counted if it is distinct, so a word repeated forty times counts once, not forty; if it is at least two signs long; and if its last sign has a reading in our table. That last rule carries the project’s oldest limit. Linear A has never been read: nobody knows what its signs sounded like. Every sound we hang on a Linear A sign is borrowed from a Linear B sign shaped the same way, and Linear B we can read. So the whole count looks through Linear B’s spectacles, and we say so out loud. One more rule, a safety bar: if the table can supply a reading for fewer than 80 of every 100 words on either side, we call no result.

The first time we ran this, in early August, the table failed that bar. On the Linear B side it could read only 69 words in every 100, under the 80 we had set. So the verdict was refused, by us, under our own rule. Not a no: a no would be a finding, and we had none, only a table too full of holes to let anyone conclude anything. We filed the refusal in the open and left it there.

Then we went back to the same published source, the sign grids compiled by the scholar John Younger, and transcribed every row we had left on the floor: 27 more signs on the Linear B side. Of those, 17 came with a reading we could use, 8 carried only readings scholars still dispute, and 2 the source itself lists with no known sound. A second person checked the new rows one at a time, against the source, before any of them touched the count.

Here is the part we knew before pressing anything. Not one of those 17 new readings ends in consonant-plus-a. Follow what that does to the count. The new rows could only add words to the countable pile, and none of the words they let in could end the way the claim is about. So on the Linear B side the share ending in consonant-plus-a could only fall, never rise. We wrote this into the frozen plan, in advance: the run could end exactly two ways. Either the table still could not see enough of Linear B and there was again no verdict, or it could see enough, and the answer came out on Owens’s side. No arrangement of the data could hand us a no. A test that cannot say no is worth running for one thing, the single thing it can still settle: whether the table now sees enough to speak at all.

It does. With the wider table the Linear B side is readable for 98 words in every 100, and the Linear A side for 90, both clear of the 80 bar. The refused verdict completes. Of the words we can score, about 33 in 100 end in consonant-plus-a in Linear A, against about 22 in 100 in Linear B. Under the way we chose to count, the lean the old claim named is there.

Every result here passes a required review before anyone reads meaning into it, run by people brought in cold, without the argument that produced it. The independent recount lands on the same figures. Its sharpest point is about that eleven-point gap, 33 against 22, and it takes most of the gap apart. The words the new rows let in, every one, end in signs whose readings never spell consonant-plus-a, and a great many of those are grammatical endings of Mycenaean Greek, the language Linear B writes. One ending does much of the work on its own: the sign read jo, which that Greek used to mark possession, close to our of, accounts by itself for 12 of every 100 countable Linear B words. Those words were always in the language; the wider table just started counting them, and they pulled the Linear B share down. Count instead only on the signs the two scripts genuinely share, like for like, and the two sit at 33 and 31. Two points, not eleven. What the review could not wash out is the direction itself: it holds at every cut we tried, at each word length, using only the most secure readings, and counting repeats instead of distinct words.

So, kept honest, this result is narrow, and it should stay narrow. What we can say: under our own way of counting, the lean the old claim described is really there, and the one thing this run added is the coverage that had been missing. What we cannot say runs longer. Nothing here reaches the language behind Linear A. The two writing systems are cousins, one grew out of the other, so a resemblance in how their words end is expected by construction and is not evidence of anything underneath. And every reading in the count still comes through Linear B’s spectacles, a borrowed lens we cannot yet take off. This result travels with its refused twin, always: the first attempt that could give no answer, and this one that can. Cite the one and you owe the other.

The last few entries built a machine that measures the plain surface of a script: how tightly its signs cluster, how long its words are, how much word variety there is. This entry is about widening the comparison. Before you can ask whether Linear A resembles some other ancient language, you have to know what is actually there to compare it against. So we went looking, and counted three candidate sources.

Two of them turned out to be the wrong kind of thing. One is a Wiktionary category of 1,602 Greek words that scholars trace back to a lost language spoken in Greece before Greek arrived: useful, but a list of dictionary headwords, not a stretch of running text. The other is eDiAna, an etymological dictionary of the minor ancient languages of Anatolia, 3,799 entries, and one where each entry already carries an editor’s analysis into roots and endings. Our census, the same fixed count we run on every script, needs running text. You cannot run it on a word list, and measuring an editor’s segmentation would measure the editor, not the language. So these two we simply counted and set aside as inventories.

The third source is a different animal. It is a corpus, the surviving inscriptions of a language taken together, and the language is Urartian. It earns its place here for a plain reason. The studies we have been auditing each compare Linear A against a single family, picked in advance by whoever wrote the study; this project measures the same things, under the same frozen rules, across several families at once, controls included. Of the three sources we added, Urartian’s is the only family that came with a real body of running text: its best-documented member is Urartian itself, while the sister language, Hurrian, has no usable machine-readable corpus we could find, so Urartian stands in for the family, and every number here carries that label. It also brings something new to the test: a completely different writing tradition, cuneiform, and a language the comparison had not seen before. If our counts were really reaching the language underneath, here was a fresh language to show it. The corpus is 816 inscriptions, 661 of them edited word by word (what scholars call lemmatised), which comes to 13,860 words of running text. As always, we wrote the rules down and froze them before looking at a single number, then measured 39 new cells, checked the arithmetic independently, and ran the whole thing twice to identical output.

The result extends the one from the previous census, and if anything it hardens. Even with the fifth script, the profiles sort by writing system, by editorial handling, and by document type, not by language. Three things show why, and each carries its own caveat.

First, the Urartian corpus looks suspiciously clean. In Linear A and Linear B the single most common mark is a placeholder for a sign too broken to read, and the published Urartian edition shows none of these at all. That can look like a tidier language. It is not. The edition has resolved each damaged sign for you. Drop to the raw text underneath and the placeholder jumps straight back to the top, 29 in every 100 signs, the same kind of hole that tops the two Aegean scripts. What looked like a fact about the language was the editing.

Second, the Urartian corpus is not one even thing. A single kind of document, royal inscriptions carved in stone, makes up about 85 in 100 of it. So what reads like the profile of Urartian is really the profile of that one monumental register. We could try removing one document type at a time to check whether the profile survives, but with a single type this dominant that check cannot pull it apart: take the royal stone away and 85 in 100 of the corpus goes with it.

Third, the pattern we flagged last time holds with a fifth script added. Line the scripts up by how tightly their commonest signs cluster, and the order is set by one thing: how many distinct signs each script uses. Urartian, with 150 signs, slots in between Linear B and Linear A, exactly where its inventory puts it. The most concentrated of all five is still an ordinary alphabet, Finnish, a living language we can read. The ranking measures the size of the sign list, not any family tie.

One more thing is worth telling, because it is the method working as intended. Partway through the run, the program stopped itself. A number we had frozen in advance was the wrong population: we had written down the count of dictionary words for Urartian, 7,990, where the measurement needed the count of words in the running text, 13,860. The two are different things (the running-text count is larger partly because it includes slots where every sign is lost), and the frozen rule had named the wrong one. We did not quietly swap the number to let the run finish. We stopped, recorded the mistake in the open as a correction, and only then continued with the right population. A frozen number is only worth freezing if you are willing to be caught by it.

So, as before, the honest form of this entry is a negative one. Adding a fifth script, in a different writing tradition and a language new to the comparison, does not make Linear A resemble anything. The census does not say Linear A is close to Urartian, or to anyone: none of these numbers is a distance, and the two scripts share not one unit you could measure a distance in. What the census does say is narrow and worth saying plainly. Measured on the surface, these scripts are told apart by how they were written, how they were edited, and what they record, not by the language underneath. It marks what does not distinguish them, never who Linear A is like.

The previous entry ended by getting the two scripts side by side in the machine. This entry is the first thing we did with them there: a census. We took Linear A and Linear B, and alongside them two languages we can already read, Sumerian and Finnish, and measured the same plain things on all four, by the same rules, written down before we looked. The two known languages are there as a ruler. If a measurement really reaches the language, it should treat the two familiar languages differently from the two Aegean scripts. If instead it treats them alike, then the measurement was never seeing the language to begin with. That is the whole design in one sentence.

We measured three things, all on the surface of the writing, none of them a reading of it. First, how tightly the signs cluster: whether a few signs do most of the work or the load is spread across many. Second, how long the words are. Third, how much word variety there is, which we compare fairly by drawing the same number of words from each script and counting how many come out different. Start with the clustering.

Here is the measurement turning on itself. If tight clustering were a fact about the language, the two known languages should not sit at the extremes. They do. The most concentrated of all four is Finnish, an ordinary alphabet of a living language, where the top twenty letters cover 97 in 100 of the text. The least concentrated is Sumerian, whose transliteration uses more than 1,500 distinct marks. Linear B, with 113 signs, and Linear A, with 292, fall in between, in that order. Line the four up by how many distinct signs each uses and the clustering drops in lockstep. So this number reports the size of the sign inventory, an accident of how each script was built and transcribed, not anything about the language it carries.

The same lesson returns on word variety. Because a short word written in a couple of syllabic signs repeats far more often than a word spelled out in letters, the raw levels cannot be compared across scripts, and we do not compare them. What we can read is where Linear A falls. Draw two thousand words from each script and count how many are different: Linear A comes out at 43 in 100, between Linear B at 27 and the two known languages at 66 and 77. It sits in the middle, not at an edge. On the separate profile of which signs open and close words, Linear A is the least sharply patterned of the four, and the most sharply patterned is the Finnish control. Wherever we look, Linear A declines to stand out.

Two things have to be said out loud, because they inflate any resemblance between the two Aegean scripts. The first is a mark we write as a star, which stands for a sign too broken to read. It is the single most common “sign” in both Linear A and Linear B, about a quarter of everything in Linear A and half of everything in Linear B. A good part of what makes the two look alike is two lists dominated by the same hole, not by the same language. The second is that Linear A and Linear B share the same family of writing and are both almost entirely bookkeeping. That makes them the pair most likely to look related when they are only built alike, and for that pair we could not even separate the results by type of document, a check we could run only on Sumerian. We have not removed these two effects. We are declaring them.

So the honest form of this entry is a negative one. Measured the same way, on the surface, these four scripts sort themselves by how they were written and what they were written for, not by the language they carry, and the two outsiders behave just like the two Aegean scripts. None of these numbers is a distance, and none of them says Linear A is close to anyone. What the phase is worth is the method: the first time the same rulers have been laid across all four at once, and a result reported in the negative instead of dressed up. The next job is the one the previous entry pointed to, widening the bridge of sign equivalences between Linear A and Linear B, so that a real comparison of those two, the one everyone wants, can at last be made able to see.

The previous entry ended with a promise: the Linear B texts come in, the locked test runs, and we report what falls out, whichever way it falls. This is that report. First, the choreography, because it is the part that gives the numbers their meaning. Every piece of code that would read the new texts, every rule that would score them, and one extra test besides, were written, reviewed and committed before the download, so that the data could not whisper to the rules. Then we fetched the tablets from the same public project our Linear A texts come from, locked to an exact version so anyone can redo the counts: 5,832 inscriptions, the tablets of Knossos, Pylos and the other Mycenaean palaces. One technical seam needed a stitch after the download, a difference in how the file wraps its records; the fix was declared in the open and checked twice by reviewers who had no stake in it, because it is exactly the kind of small quiet change that a project like this has to keep visible.

The name test works like this. In Linear B accounting, a name tends to open a line that ends in a commodity sign and a number: so-and-so, sheep, a hundred. We look for words in that position, but only in documents that have at least three such lines, the written-down rule for “this is a list of somebodies”. That rule keeps 128 documents out of 5,832, and in them 472 distinct words sitting where names sit. To compare them with our 74 Linear A candidates at all, each Linear B word must be carried across to the other script over a bridge table of sign equivalences, itself a hypothesis, declared and versioned, and the bridge carries only about half of them: 233 of the 472. Against those, our candidates of three and four signs, 41 words, the lengths where a match would be hardest to dismiss.

The result is zero matches at those lengths. Two two-sign coincidences exist, and two signs is short enough that we report them as bookkeeping, not as evidence. Now the honest reading of that zero, which the mandatory review every result passes before we interpret it has spelled out. The comparison against randomly shuffled data, run ten thousand times, also produced zero essentially every time: a test where even chance cannot score is a test with no power to see, and we say so rather than dress the zero as a discovery about the two languages. The check on pa-i-to makes it concrete: the word is in the tablets 47 times, and it is widely identified as Phaistos, the very town whose ruins sit near where many Linear A tablets were found; ten of those occurrences open a line, two even open name-shaped lines, but in documents with only two such lines, one short of our three. The rule was written for long list tablets, and Linear B keeps its place names mostly on short ones. On top of that, a blind check we had also declared in advance: fifty of the 472 slot words were shown to a reviewer who did not know our candidate list, with the question, is this an identified name in the scholarship? Thirteen of fifty were. Under the bar we had set before looking, that is too few, and the whole exercise steps down from a test to a census. So the entry’s first result is a boundary, not a verdict: this sieve, built this way, cannot yet see what it was built to look for. What it cost us is real and declared; what it did not cost us is a false claim either way.

The second test came from the audit series. A scholar of Aegean scripts, Gareth Owens, wrote in passing that more sign groups end in consonant-plus-a in Linear A than in Linear B, one line in a 1990s article, never backed by counts. It is exactly the kind of claim a corpus can check, so when the audit flagged it we froze a way to count it, on both scripts at once, before either side had been counted. With the Linear B texts in, both sides ran in the same hour. The numbers lean his way: of the words we can score, 33 in 100 end that way in Linear A against 31 in 100 in Linear B, and the gap widens when we restrict to the signs whose readings are most secure. But the same frozen recipe included a safety bar: if the sign table lets us score less than 80 in 100 of a script’s words, the comparison does not count. Linear B came in at 69. So the verdict is withheld, by us, under our own rule, and the lean stays what it is: a lean, on the record, waiting for a wider bridge.

It is worth pausing on what kind of night this was. Rules written before looking are usually praised for stopping you from claiming too much when the result flatters you. Tonight they did something less discussed: they stopped us from concluding anything at all, twice, when the instruments turned out to be too blunt, and they forced the bluntness itself into the open, in numbers. Both refusals are published in full in the project’s public list of declared tests, alongside the plans they enforced. And both point at the same next step. The bottleneck of the night was the bridge table between the scripts, 61 sign equivalences that carry only half of what we need; widening it, sign by sign, with each addition argued and versioned before any re-run, is now the declared first problem of the project’s next phase, the first symmetric comparison of the two corpora. The two scripts are finally side by side in the machine. Making the comparison able to see is the work.

For three entries we did nothing but audit: we took other people’s published claims about the language behind Linear A, the undeciphered script of Minoan Crete, and asked which of them could be checked at all. This entry changes direction. It is the first small piece we build ourselves, and it starts from the safest end of a treacherous field. To read one word of Linear A you would need the language, and nobody has it. To notice that a word behaves like a name, you do not need to read anything.

Here is why behaviour is enough. Of the 1,721 inscriptions that survive, most are bureaucracy: clay tablets that record goods, people and quantities. The standard line of such a document is a word followed by a number, sometimes with a small commodity sign between them. We call a document a list when it has at least three lines of that shape, and 171 documents qualify. If Minoan scribes recorded who delivered what, and where it came from, then the words for people and places should sit in predictable places: at the head of a list, or in the rows, next to the numbers, and the same ones should return on other documents.

We turned that expectation into three rules and froze them, meaning we wrote them into our public list of declared tests before running anything, so we could not nudge them after seeing the result. A candidate must be at least two signs long. It must occur in a list position, or open a list document. And it must appear in at least two different documents. No meanings, no sound values, no outside lists of names: the rules see only where a word sits inside the inscriptions. One limit is declared rather than hidden: about 40 in 100 word occurrences are too damaged to identify, and we never guess a damaged word into a reading, so the census works on roughly six occurrences out of ten. Of the 618 distinct words that are fully readable and made of ordinary syllable signs, 74 pass all three rules.

Then comes the part where honesty costs something. One other project, as far as we know the only one, works on this same question in the same testable spirit: it is published under a pseudonym on GitHub (the DecipherLinearA repository, archived with a DOI, cited in full on our sources page), and we cite it the way we would cite any colleague, without contacting anyone and without guessing who is behind it. It approaches the problem from the opposite shore. Linear B, the later script of Mycenaean Greece, can be read, and its tablets carry personal names and place names in well-understood positions. That project’s nine published anchor words, as it calls them, are Linear A words whose written shape also appears in Linear B where names go. Nine public words make a natural check for our list, and we ran it, but on one condition stated up front: since we had read their nine before our rules ever ran, nothing in the comparison may be called a discovery, and no statistics are attached to it. Their selection rules, on the other hand, we have never opened, so our three rules were not written with theirs in view.

The comparison came back like this. All nine of their words exist among our 618. Four of them, da-i-pi-ta, pa-i-to, pa-ra-ne and ta-na-ti, also pass our three rules and sit inside the 74. Five do not. One of the five, ki-da-ro, misses a single rule: it never returns across two documents. The other four miss the list-position rule and the recurrence rule both.

Two sieves were held over the same nine words, and they kept the same four. That sentence is the whole finding, and it needs spelling out what it is not. It does not say the five outside are not names. Our rules demand a kind of prominence inside the Linear A bookkeeping that their criterion, a matching written shape across two scripts, never asked for; a word can fail our sieve and pass theirs without either sieve being wrong. It does not say we confirmed four of their names: appearing on both lists is not a reading, and we attach no percentage and no statistical weight to four out of nine. A comparison whose answer was partly known in advance can be reported, listed and learned from; it cannot be scored.

The part of this entry we care most about has not happened yet. The real question is whether our 74, chosen by behaviour alone, show up where names go in Linear B beyond the nine anchors everyone already knows. That test is fully written: what counts as a match, what counts as a name position, what thresholds apply, and the rule that the verdict is counted only on new matches, with the nine known words set aside. It is locked in our public list of declared tests. And the Linear B texts themselves are not in the project: we froze the test before bringing in the data it needs, so that the data could not whisper to the rules while they were being written. One caution stays attached even here. Linear B has been read for over seventy years, and these are the same clay tablets scholars have always studied, not fresh soil; freezing first is the strongest protection available in a field with no unseen data, and we present it as that, not as a guarantee.

The design of all this went through the same hostile review as every design here, and it did not survive intact. An outside reviewer, paid a few cents to break the plan, broke it in nine places, and the most useful catch was about the freeze itself. An early draft declared that no one on this project had ever touched Linear B data. Written that way, it was false: during scouting, an assistant had downloaded one Linear B file to check its format, keeping the shape of the data and none of its words. The design now records that access in the open, and states the rule that protects the test from here on: from the moment of the freeze, zero access to Linear B data until the day the test runs. We publish the correction along with the plan because a protection nobody checks is a promise, not a protection. The next step belongs to the data: the Linear B texts come in, the locked test runs, and we will report what falls out, whichever way it falls.

The stone was found in 1994 at Kaphkania, three kilometres from ancient Olympia: egg-shaped, five centimetres, 48 grams, nine signs in two lines. The layer it came from is dated to the seventeenth century BC. If it is genuinely that old, and genuinely mainland writing, it rewrites a chapter of history, because the earliest secure writing from the Greek mainland, the Linear B archives, comes centuries later. That is exactly how one senior scholar announced it.

The four-page article we audited, from 1998/99, pushed back: could the pebble be Minoan work instead, the product of the older Cretan world? Its conclusion is worded as a question, with two “perhaps” in it, and its own footnote records, wryly, that the stone was allegedly found on April the first. We did not try to settle who is right about the pebble. Our audit asks the same narrow pair of questions we asked of the other two texts: which of its statements could ever be shown wrong, and which of those can we check with what we hold?

Most of it, once again, is other people’s territory: Mycenaean name lists, a name that also occurs in Homer, the votive customs of Bronze Age Crete. And the pebble itself is not in our data: our material is the 1,721 surviving Linear A inscriptions, the writing of Minoan Crete, and not one of them is from Olympia. On paper, this was the audit where we could touch the least.

But the article leans on details from the Cretan material we do hold, and three of them we could look up directly. First, it compares one of the pebble’s words to two Minoan words from the site of Haghia Triada. One of the two, KA-RO-NA, sits exactly where the article says, on tablet HT11a from Haghia Triada. The other, KA-RO-PA₂, we cannot find anywhere in our 1,721 inscriptions; the closest candidate is a word on tablet HT75 whose final sign is lost to damage. Editions of these texts differ in how they read worn signs, so we record “not found in our data” and leave it there. Not found is not the same as wrong.

Second, the whole quarrel turns on three of the pebble’s nine signs, the ones read as SO, QO and JO. Both sides of the dispute start from the same premise: those three signs had been seen only in Linear B, the later mainland script, never in Linear A. Our count agrees. Across all 1,721 inscriptions the three signs never appear, not once, while the other five signs of the pebble are ordinary Linear A signs with 45 to 285 appearances each. The premise holds in our data; what it proves about the pebble is exactly what the two sides have been arguing about for thirty years, and that argument is not ours to settle.

One of the article’s arguments we would have liked to test is a counting claim: that more sign groups end in a consonant-plus-A syllable in Linear A than in Linear B. Half of that count we could run today, on our side of the fence. The other half needs a Linear B corpus, which this project has never brought in. Half a test is worse than no test, so we wrote the test down in our public list of declared tests and parked it until the day a Linear B corpus enters the project.

That closes the series we announced: three published texts, one per family of ideas about the Minoan language, each read statement by statement with the same two questions. The shape repeated every time. At the centre of each text sits a detail you can check by opening the books; around it, an inference that needs a specialist’s judgment. The checking part is ours, and we did it: some details held, some did not, most of the rest belongs to fields we do not practise. What comes next is different. Instead of auditing other people’s proposals, we start on a small piece of our own, from the safest end: looking for the Minoan words that behave like names of people and places, without claiming to read any of them.

Linear A is the older Minoan script, and nobody can read it: we can copy the signs but we do not know the language. “Semitic” is the family that includes Phoenician, Hebrew and others. Gordon’s 1967 article set out to show that Minoan belongs to it. It is a famous minority position, the kind of thing people still search for.

We are not judging whether he was right. We cannot: the script is undeciphered. What we can do is sort his claims by a plainer question. Are they the sort of thing that could ever be shown wrong, and if so, could we be the ones to do it? Almost every claim he makes is a comparison: this Minoan word matches that Phoenician or Ugaritic word. Comparisons like that can be checked, but by someone who works on ancient Semitic languages, going back to those texts. We have the Linear A inscriptions and a way of counting signs, nothing more. So six of his eleven claims fall entirely outside what we can do, and the rest we can only partly inspect.

One claim was different. Gordon said the same Minoan word turns up with two different beginnings, “ya-” and “a-”, as if the two were interchangeable, and he gave an example word to prove it. That is a pattern we can look for: we can search every surviving inscription for words that are identical except for that first sign. So we wrote the test rules down in advance, then ran them across all 1,721 inscriptions.

The pattern is there, barely. Only two words out of dozens show both beginnings, and both are short and appear just once each, which means they could easily be two different words that happen to start alike. And Gordon’s own example word does not survive the test in our data, for a dull reason: the one form that would have matched sits on a broken tablet, and our rules set damaged words aside before we looked. So the honest verdict is a small point in his favour, not proof, and not a strike against him either. A specialist’s own review of Gordon agrees with him on this exact point; our count simply does not reproduce it.

The wider result is the quieter one. On the single claim of his we could put to the test, the answer came back weak. On the other ten, the honest thing to say is that they are someone else’s job, not ours.

The previous entry told how we tried, and failed, to build a mechanical rulebook for judging published claims about the language behind Linear A. The fallback we announced was a reasoned reading: arguments, openly ours. While preparing those readings we found that one statement on our material's central list allows something firmer than an argument. The list, thirteen printed statements from a scholarly article of the early 1990s, concerns the language’s endings, pronouns and small connecting words; its thirteenth statement says that those small words appear in a fixed order. The claim lives in a specific place: the dedication formula, a recurring text that Bronze Age Crete carved and painted on stone offering vessels, of which our copy of the surviving inscriptions holds 165 examples. A fixed order is a rare luxury for a tester. It does not need any theory to be checked: it needs one counterexample to die, and a closed set of inscriptions to hunt for it in.

Grading your own exam is still the trap, so before looking we froze the rules, meaning we wrote them down and committed them, timestamped, before running anything. Which inscriptions count: all 165 in the vessel series, no cherry-picking. Which signs count as the five little words: fixed by catalogue number, among them the sign numbered 301, which has no agreed sound value and so can hide behind no reading. What counts as a chain: two or more of the five side by side inside one word, 31 of which turned up, from 26 different inscriptions. And one deliberately uncomfortable choice: no filter deciding which chains “really” belong to the formula, because sorting chains by hand, after seeing their order, is exactly the back door a tester must brick up. One more honesty note: the article states that the order is fixed but never prints the order itself; we rebuilt it from the article’s own cited examples, so the target of the test is the claim as we reconstructed it.

By the frozen rules, the claim fails. Five pairs of the little words occur in both orders somewhere in the material, and one reversed pair is all that “always” can afford. That is the verdict, it was fixed before we looked, and it stands.

Then we handed the result to two fresh reviewers, people uninvolved in building the test, for the mandatory review every result passes before we interpret it. They disagreed on emphasis and agreed on the map. In all 17 chains that contain the sign 301, the stretch from 301 onward runs in the same order. Every reverse arrow traces back to three places at the edge of the material: an inscription broken on both sides, where the offending word sits right at the break; the inked cup from Knossos, whose text is not the dedication formula; and words that open with the syllable ja. It is tempting to declare all three irrelevant and the order confirmed. Both reviewers, from opposite starting points, refused: deciding now that those chains do not count is the same after-the-fact sorting our rules forbade, and one dedication with the full core carries a word-initial ja itself, so splitting the two kinds of ja is a choice, not a fact. The two lines we are entitled to are these. As stated, on the whole declared material, the fixed order is falsified. The falsification never touches the 17-chain core, which is as uniform as the claim wanted.

What this does not mean: that the article’s larger thesis is wrong, or that its author was careless. The order we tested is our reconstruction; the failure sits at the edges of an all-inclusive net that the author never promised to cast; and a narrower, cleaner test of the core alone could still be declared in advance and run. We have not run it, and until someone does, “the core is confirmed” may not be said. What it does mean is a first for this project: after sixteen entries of measuring quantities related to published claims, this is the first time one statement from a published list has been tested directly, rules first, full evidence printed above, verdict included. It came out a no. Publishing the no, with the map that makes it interesting, is the whole point of working this way.

Over the last century scholars have proposed many identities for the language written in Linear A: a relative of Greek, a Semitic language, a member of the Anatolian family of ancient Turkey, and more. The next stage of this project is to read some of those proposals closely and ask a question that comes before “is this right?”: what would it take to prove each claim wrong, and who could do it? A claim that no possible observation could contradict is a different kind of statement from one that sticks its neck out, and a reader deserves to know which kind is in front of them.

There is an obvious trap in that plan. We would be both the examiners and an interested party, and nothing slants a verdict like grading your own exam. So before touching anyone’s work we tried to build the instrument first: a rulebook, meaning a fixed list of questions to ask of each claim, the answers allowed for each question, and a decision table that turns the answers into a verdict with no step left to taste. The test for such a rulebook is easy to state. Hand it, together with an author’s exact sentences, to two readers who have never met. If the rules are tight, both must land on the same verdict. If they can land on different ones, the rulebook is not measuring the claim; it is measuring the reader.

As material we used the central list of a scholarly article from the early 1990s which argues that the language of the inscriptions belongs to a specific ancient language family: thirteen short printed statements about the language’s endings, pronouns, small connecting words and word order. We copied each statement letter by letter from photographs of the printed pages, page numbers attached, after discovering that the automatic text layer of our PDF garbles exactly the abbreviations that matter. Three of the thirteen served as the bench test, and because two of them name two alternative forms at once, from the third draft onward they were handled as five separate pieces. Every round used a fresh reviewer: someone who had no part in writing the rules, was handed only the rulebook and the sentences, and had to answer one question. Following the rules to the letter, does each piece get exactly one verdict?

It never did. The first draft failed before the question could even be asked: four reviews out of four found that it had questions but no rule leading from the answers to a verdict. The second draft added the decision table, and its reviewer showed that on the three bench sentences a strict reader could still reach five, five and three different verdicts. We fixed every defect that report listed. The third draft’s reviewer found the counts lower, but the unwritten choice had moved into the step before the rules even start: how a sentence splits into pieces, and which of the author’s cited examples belongs to which piece. The fourth draft closed that, and one piece reached a single verdict for the first time; the choice moved again, into the test that decides what kind of sentence you are reading. The fifth draft finally closed the piece that had resisted three rounds. It broke anyway. Of its two new breaking defects, one sat in a field that a previous fix had made mandatory, and the other had been opened by the wording of a previous fix itself. The four review reports list forty-one numbered defects in all.

Lined up, the five reports show one pattern, and it is the real finding of this entry. The free choice, the point where a reader must decide something the rules do not decide for them, never disappeared. Each fix closed the door it was aimed at, and the choice reappeared one step earlier in the chain: from the verdict, into a field, into the preparation of the material, into the test for a field. The reviewer of the last round wrote that the remaining defect could be closed with one more line of text, and that they knew how to write it. The reviewer of the round before had said almost the same thing. That is precisely how this kind of work swallows people: the next fix always looks cheap.

Which is why, before the last round, we had written down a stopping rule: if the fourth re-test still finds a breaking defect, there is no fifth, and the exercise is scaled down. Rules like that exist for the moment when the work is nearly good: each round had made the rulebook better, and that is what makes stopping hard. The rule fired. We stopped.

Two things this failure does not mean. It does not mean the published claims are wrong or badly made: nothing in this entry judges any of them, and every defect in those forty-one findings was ours, in our rulebook, not in the article we used as material. And it does not mean the claims cannot be examined. It means we cannot honestly examine them with a machine-like instrument that removes our judgment, at least not on this material: single printed sentences about a language nobody can read. So the next entries change shape. Instead of a scoreboard, a reasoned reading of each claim, quote by quote with page numbers, where the argument is openly ours and open to attack. We promised at the start of this project that our own failures get published first and with the most space. This is one, and it tells you exactly how much to trust what follows: as an argument, not as a measurement.

On the evening of July 3, 1908, at Phaistos in southern Crete, a clay disc turned up in a room of a building beside the Minoan palace. The excavation was Italian, directed by Luigi Pernier, but the find itself was made during an evening inspection by the local foreman, outside the regular digging. Hold on to that detail; it returns near the end of this entry. The disc now sits in the Heraklion museum. Both faces carry signs pressed one by one into the wet clay with individual stamps, one punch per sign: 45 distinct signs, 241 impressions in all by the classic count, 242 by a recent one. The disc was then fired on purpose, unlike the Linear A tablets, which survive only because palace fires baked them by accident. Someone made this object to last.

It is not Linear A. Bronze-Age Crete used three related scripts, Cretan Hieroglyphic, Linear A and Linear B, and the Disc’s signs match none of them. A Linear A tablet lay a short distance away in the same room, which shows the two systems lived side by side and nothing more. One connection does survive scrutiny, and it concerns the language rather than the writing: statistical comparisons published in 2000 and 2018 measured how signs repeat and vary on the Disc against Linear A and Linear B texts of the same length, and both found the Disc behaving like Linear A and unlike Linear B. The language underneath may be the one our corpus records. The script is not, so the Disc stays out of our experiments entirely: our whole method needs a corpus, and this object has none.

Even its date is an interval, not a year. No laboratory has ever dated the object itself; everything rests on the layer of earth it lay in, and that layer had been disturbed in later antiquity. The published positions run from a cautious window of roughly 1850 to 1600 BC to a recent argument, based on the pottery found around it, for about 1750 BC. Proposals as far apart as 2100 and 1100 BC exist in the literature. The museum label prints “c. 1700-1650 BC”; the label is more confident than the evidence under it.

Because every sign was pressed with a ready-made punch, the Disc is often called the earliest printed text. The epigrapher Yves Duhoux put that away in one dry sentence: “It is not, as one often reads, the earliest example of printing but rather of a kind of typewriting.” Printing means running off copies. These punches, as far as anyone can show, were used for this one object and never again, and the text was even corrected after stamping, which no print run does. Somebody owned a beautiful set of tools and left us a single page.

What the Disc does collect is decipherments. Duhoux again, in a peer-reviewed journal: “Suggested decipherments are legion, and the propositions go in all possible and imaginary directions: Basque, Chinese (!), Dravidian, Greek, Hittite, Luwian, ‘Pelasgian,’ Semitic, Slavic, and Sumerian, to name a few.” A 2008 catalog of attempts and commentary runs past a hundred entries, and its own compiler called it incomplete two months later. Duhoux also reports that John Chadwick, the man who helped read Linear B, asked near the end of his life that people “kindly not send” him their solutions. The readings on offer include an adventure story, a prayer, a calendar, a court list, a board game and sheet music. If one of these decipherments is right, all the others are wrong. And here is the part that matters for this site: not one of them can be shown to be wrong. A reading built to fit the only text there is will fit it. No second document exists to embarrass it.

That is what the chart above is for. Our whole project lives inside 1,721 inscriptions carrying 5,947 readable sign occurrences, and we still spend most of this log saying “not enough data”: too few examples of a word, too few tablets outside one archive. The Disc is that problem taken to its limit. Statistics lives on repetition: you learn what a sign does by catching it in different company, and one document gives you one company. It has even been suggested in the scholarly literature that the signs may not encode language at all; the article sits behind a paywall we have not crossed, so we report the suggestion and nothing more. With a corpus of one, even “is this writing?” stays open.

Whether the Disc could ever be read, even in principle, turns out to be a statistics dispute too. Code-breakers use a rule called unicity distance: a coded text shorter than a certain threshold, which depends on how many signs the writing system has, cannot have a unique, checkable solution. Assume the full script had about 100 signs, as one classic handbook did, and the threshold lands near 225 while the Disc carries 241: too close, and that handbook concluded the text is too short to decipher. Argue instead, as Duhoux does from a statistical formula, that the full script had about 60 signs, and a checkable reading stops being impossible. The two camps disagree about decipherability because they disagree about a number: how many signs the system had beyond the 45 that happen to be stamped here. From one document, that number cannot be measured.

Then there is the question nobody has been allowed to test: whether the Disc is genuine at all. It has been called a hoax more than once, most loudly in 2008 by an antiquities-magazine editor who accused Pernier himself, publishing the accusation in his own magazine and, as he acknowledged afterwards, without peer review. Mainstream scholarship considers the Disc genuine, and the strongest argument is chronological rather than stylistic: one of its signs, the “comb”, matches marks on Phaistos objects that were excavated in 1955 and 1965. A forger working in 1908 cannot copy a find made half a century later. The clean way to settle it would be a laboratory test of when the clay was fired. It has been requested and never granted; the museum’s written reply, as quoted by the man who asked, says the disc “because of its uniqueness is considered as non movable”. The object is too singular to be tested. That sentence could be the epigraph of this entry.

We told this story because the Disc is the mirror our project looks into. Nearly everything this log has published is a “no” or a “not measurable”, and every one of those answers was possible for one reason: 1,721 documents are enough to contradict a wrong idea. The Disc contradicts nobody, which is why a century of readings has settled nothing. A corpus of one is not a harder puzzle. It is the end of puzzles: past that line there is no way left to be wrong, and a claim that cannot be wrong has nowhere to go.

Entry #013 closed our long look at the shapes of the signs. This entry turns to the other half of the tablets: the accounting. Some signs do not spell sounds but stand for a thing being counted, what scholars call a commodity sign or logogram, with standard identifications like “grain” or “wine”. They are usually followed by a number. Back in the word-counting work of entry #011 we had written down, in advance, two questions about them: do different goods carry different ranges of amounts, and do the scribes favor round numbers?

Writing a test down in advance matters because a test whose outcome you already know proves nothing: even in good faith, you can shape it until it says what you expect. And that is the trap we walked into. To design the test we measured the material, and the measuring showed us the answers: the typical amounts behind each commodity sign, and even the tally of round numbers, which one of our own reviewers computed while checking that the test made mathematical sense. The corpus of Linear A cannot grow. Once you have seen the answer on the only data there will ever be, that question is burned for you. We keep a public record of exactly what was seen, by whom, and when, and both questions are now closed for us as tests.

So this entry reports a census: a full count, no verdict, published so that anyone with cleaner hands than ours can build a fair test on top of it. The core of it: on the tablet lines, a sign of this kind sits immediately before a whole number with a readable value 287 times, across 90 distinct signs. That 287 is not a prediction that came true; it is the same count produced by two independent programs from the same frozen data, which is reassuring about our arithmetic and nothing more.

The first thing the count says is that there is no clean boundary around “goods signs”. The standard sign list marks 28 of them. But 105 compound signs, ligatures where two signs are written as one, behave exactly the same way on the tablets, and a handful of plain sound-signs do too: one of them, A304, is followed by a number 22 times out of 22. The chart shows the two populations overlapping in every band. We publish the whole spectrum and refuse to draw a line through it, because any line drawn now would be drawn after seeing the numbers, which is the same trap as above wearing different clothes.

The second thing: each sign carries its own scale. Half of the amounts after the “grain” sign are 20 or more, and the largest is 976; after “figs”, the typical amount is 2 or 3. These are raw counts, not a comparison test: the differences are not adjusted for where the tablets were found or what kind of document they are, so we do not get to say that grain “really” comes in bigger batches. They also cover only amounts written as whole numbers. Fractions exist, 743 of the 1,718 number tokens carry fraction marks, and their values are proposed by published scholarship rather than asserted by our frozen data: under that clearly labeled what-if, 47 of the 287 amounts would grow, and for 4 of them the what-if has no answer at all.

We also tried to check our raw material against an outside source, and we owe the result in plain words: the check partly failed. Our line order comes from a view of the corpus we built and audited in entry #012's phase. The outside source, the SigLA database, records the drawings of signs but does not record the numbers as signs at all. So “is this sign really followed by that number” cannot be verified there for a single one of the 287 pairs. What we could check is weaker: whether the sign itself sits among the same neighbours. On 20 sampled pairs: 11 matched, 3 disagreed on what sign is written, 1 differed only in naming convention, 1 could not be aligned, and 4 were simply absent from the source, which means “we do not know”, not “no”. Two of the fraction what-if cases sit among the disagreements, and we flag them where they appear. A visual check against the published photographs remains possible; we have deliberately not opened them, and if we ever do, the rules of that check will be written down first, for the reason this whole entry keeps repeating.

One more thing the census has to say out loud: seven of every ten sign-and-number pairs come from a single site, the Haghia Triada archive. And the skew runs twice, because readability is not even either: at Haghia Triada about 15% of the isolated signs are too damaged to read, at Khania it is 48%. What survived, and what survived legibly, both lean the same way, so every count above is first of all a count of one warehouse.

This is the least glamorous entry of the log so far, and probably the most honest one. The two questions we burned stay burned. What replaces them is on the record: every count above can be regenerated from the frozen data with one script, the full spectrum of 220 sign types is published beside it, and the whole thing now waits for a test that someone can still run fairly.

Phase 5a asks a mechanical question, not a linguistic one. The standard list of Linear A signs is fixed: each sign has a number. Two things could be wrong with such a list. A sign counted as one might really be two that scholars merged, or two signs on the list might really be one. This has nothing to do with what the signs sounded like. It is only about whether the shapes pressed into the clay, sorted by a computer that was never told the list, agree with the list drawn from them.

The material is 3,174 sign-pictures cut from the SigLA database, one for each time a sign appears. We tested only the 43 signs that turn up at least twenty times, which is 2,836 of those pictures; the other 89 signs are too rare to say anything about, and we leave them out loud. So everything below is about 43 of the 132 signs, not the whole script. The pictures stay on our machine; only counts and figures are published.

One rule was fixed before anything ran. We measure each sign-picture two ways at once: an old, plain method that simply tallies which way the edges run, and a modern neural network. The rule: if the network is not clearly better than the plain method at recovering the known sign list, we trust the plain one. It was not better. On a calibration check the plain method matched the list more closely than the network did, and a second plain method we added during review did worse than both. So the plain method leads, chosen by a rule written before we had seen a single result.

On that leading method the answer is a flat zero. Not one of the 43 signs splits in two. Every one of the 903 possible pairs stays apart, so nothing merges. Stop there and the headline writes itself: computer vision confirms the sign list. That headline is wrong three times over.

First, the zero on splitting comes from where we drew a line, not from the statistics. The safety test we set up to catch false splits turned out to be weak, which if anything should have made splits easier to declare. None appeared, because none was sharp enough to clear a fixed bar we set in advance. The closest sign, AB31, fell short of that bar by one part in a hundred. So the honest sentence is “nothing was sharp enough to call,” not “the statistics ruled it out.”

Second, “every pair stays apart” rests on something almost automatic. When you describe each picture with more than a thousand numbers, almost any two labelled piles can be told apart; that is a quirk of having so many numbers, not proof the signs are truly different. Our own data admits it: on 63 of the 903 pairs, the computer cannot actually find the line between the two signs on its own. The easiest pair to confuse, AB55 and AB56, sits right at the edge. The honest claim is small: no pair fell past the merge line we set in advance, and the nearest one came close. Not “the sign list is proven right.”

Third, and this is the part worth keeping: how many splits you find depends heavily on which method you look with. Zero with the leading method, seven with the second plain one, twenty-three with the network. Five signs show up in the overlap, splitting under two methods that fail in unrelated ways: AB03, AB27, AB31, AB37, AB77. They are not an accident of which site or which scribe made them; we checked, and the split holds inside a single findspot. But they fail the leading method, and two of them are the same near-misses that fell one part in a hundred short. We will not promote them, because that would mean overruling our own rule after seeing the answer. We will not dismiss them either, because the obvious way to explain them away, plain differences in size and ink, is the one thing our fixed test never checked.

Before writing this we wrote that test down in advance, and we have now run it: do those five splits survive once you account for how big each sign is and how dark its marks are? Four of the five faded. Two of them, AB37 and AB77, split on nothing but those surface differences. Two more, AB03 and AB31, held up under only one of the two other methods, not both. One sign, AB27, kept a genuine split even after the control. But even AB27 does not clear the bar of the method we trust, and it cannot count as a fix to the list: the same database drew both the pictures we sorted and the labels we checked them against, so nothing here can confirm that list from the outside. So the leading method’s zero mostly held, and a single sign is left as a thin thread for someone to pull later, not a result.

The mandatory review, the one every result of ours passes before we interpret it, did its job again. It turned a tidy zero into a “no” we can stand behind for most of the list, a smaller claim about the pairs, and five signs sent through a test we wrote down in advance instead of a story we told after, which left just one of them standing. No sign got a sound. Nothing here touches what language lies underneath. That bar has not moved.

This morning we wrote nine tests down in the ledger, the public list where we declare what we will measure before running it, then ran all of Phase 4 in one day: word structure on one side, tablet arithmetic on the other. One rule is new and worth stating up front: from now on, every headline claim on this site names the group of tests, declared in advance, that it stands on, so that opening new groups each phase can never quietly reset the odds in our favor. Today’s headline stands on the group whose code in the ledger is F10, and it is a no.

First, the machine. We wrote a small compressor that reads the 647 complete words of two signs or more and tries to explain them as reusable parts, the way “un-help-ful” explains English. It works: it found 226 recurring pieces, cut the description of the vocabulary by more than a third, and produced a tidy list of suffix-like endings. Fifty-nine word-stems even take two or more different endings, which is what paradigms look like. If we stopped here, this would be a lovely paragraph about Linear A morphology.

We did not stop here. The design of the test, put in writing before the run, demands that each of the two measures beat two randomly scrambled versions of the corpus, the body of inscriptions, at once. Scramble one: spin each word like a dial, so A‑B‑C becomes B‑C‑A, keeping its signs and length but moving its edges. Scramble two: swap signs between words while keeping each sign’s favorite position. The compression gain died on the first scramble. Spun words compress just as well as real ones, which means the gain was never about word edges at all; it lives in sign frequencies and local sign habits we had already measured in Phase 2. The paradigm count beat neither cleanly. Verdict, with the label we promised: Fact under test group F10, declared before we saw the data, this corpus shows no word-internal structure beyond what its known sign habits explain. Anyone who wants to claim Linear A suffixes now has these two scrambles to beat. We are leaving them armed.

We knew the deck was stacked: 72 percent of the vocabulary occurs exactly once, and one-off words carry no repetition for a test to grip. That number was in #011 before any of this ran. A no measured against declared odds is still information, and it is the kind this project exists to produce.

The arithmetic half went differently. Using the line structure we recovered from our source (a by-product worth its own entry: the corpus encodes bookkeeping entries, one word with its quantity per line), we walked every KU‑RO total on the tablets and added up the entries above it, by machine, all thirty. Fourteen segments are fully checkable as plain integers. Six of those fourteen add up exactly.

The eight misses are where it gets good. Two of them stop being misses the moment you read the tablet’s structure instead of our frozen rule: on HT 88 the total counts only the six entries after the KI‑RO line, exactly; on HT 127b the second total is cumulative, the five entries plus the previous total, 136 + 156 = 292, exactly. Two more lose their entries to surface damage. What remains are three near misses of one to four units, the honest texture of a working ledger kept by humans, and one open case off by twenty. And under the values a Bologna research group has proposed for the fraction signs, adopted here as a labeled hypothesis rather than a fact, tablet HT 104 balances to the digit: 95 = 95.

That Bologna paper, the one solid decoding result in this field in a decade, is also the one we set out to replicate today. Straight answer: we could not regenerate their solution counts from the published text alone; the paper describes its constraints in prose, the solver model itself is unpublished, and under every reading we tried the numbers land elsewhere. We wrote that down and stopped, because bending constraints until totals match is how you fabricate a replication. What we could test independently held: their proposed values satisfy every stated constraint, and their key statistical signature, recomputed on our own recount of every fraction sign in our corpus, lands inside their own published window. The values survive contact with a second dataset. The derivation, for now, remains theirs alone.

Phase 4 closed in a day: one honest no with teeth, a ledger audited to the digit, and a replication that reports exactly how far it got.

Phase 3, our first test on the sounds of the signs, ended in entry #010 with a no we could stand behind, and a lesson we kept writing down: the evidence base was 48 words when we thought it was 233. Phase 4 is about word structure and about the numbers on the tablets, so before we put a single test in writing we did the boring thing first. We counted the pile. What follows is the shape of the material, measured so the tests we design next fit the thing they are testing and not a thing we imagined.

Start with the words themselves. Of the 1,474 distinct words in the corpus, 839 are complete: every sign readable, no break, no gap. Sorted by length they pile up where you would expect a bookkeeping script to pile up, at two and three signs, and thin out fast. Forty-seven types reach five signs. Past seven there are seven words in the whole corpus, and the single longest, at nineteen signs, is inscribed on a metal object from Knossos, not on anything anyone was balancing in a ledger.

Then the number that governs everything else. Of those 839 complete words, 607 appear exactly once. Seventy-two percent of the vocabulary is spoken a single time. That is not a defect in our reading; it is what a corpus of names and one-off entries looks like. It is also the Phase 3 lesson in general form: a statistic about word structure has far less to stand on than a raw word count suggests, because most of the words carry no repetition to measure. We would rather know that before we design the test than discover it in the final review.

The tablets are not all fog. In one corner they behave like arithmetic. The word KU-RO, long read as a running total, appears thirty times, and in all thirty it is followed on the same line by a number. Not most of the time. Every time. Its companions behave too: KI-RO, plausibly a deficit or amount owed, thirteen times, and PO-TO-KU-RO, a grand total, twice. This is the seam Phase 4’s metrology half will pull on, because a word that always sits beside a number is a word whose job you can test without knowing its sounds.

Last, an honest inventory of what we do not have yet. Photographs of individual signs exist: a scholarly database carries 5,137 labeled cutouts, one per sign occurrence, which we can study directly. A graded map of how each Linear A sign relates to its Linear B descendant, the kind of thing a machine could read, does not exist; the fine version lives only in a 2020 book and its appendix, which is a trip to a library, not a download. And of five external comparisons a decipherment usually reaches for, only some are open to us.

None of this moves the needle on reading Linear A, and none of it is meant to. It is the drawer, counted and labeled, before we pick a tool out of it.

Phase 3 asks the narrow version of the oldest question here. Everyone reads Linear A signs with Linear B sound values; it is the borrowed hypothesis every decipherment attempt starts from, including the one we measured in #008. We cannot test whether those values are right. What we can test is whether this particular sign-to-sound table produces structure that alternatives would not. So the rule, put in writing before any run: a pattern in the sounds counts only if it beats two comparisons against chance at once. Comparison one: shuffle which sign carries which sound, swapping only values compatible with each other, and the pattern should die. Comparison two: shuffle the order of signs inside each word while keeping the true sounds, and it should die there too. Beat only the second and you have rediscovered, in a costume, the position preferences of signs we had already measured in Phase 2. Beat both, and the table itself is doing the work.

The setup: John Younger’s correspondence grid, the table in which a senior Linear A scholar proposes, for each sign, the sound it should carry. Locked to a dated copy, read by a mechanical rule with no judgment calls. A cell with no question mark counts as secure. That gave fifteen secure signs, 233 usable words, fourteen consonant and vowel features, ten thousand reshuffles per comparison.

One caveat a specialist will have spotted two paragraphs ago: Younger’s values exist because the same words appear in both scripts, some of them in this very corpus. Part of the positional behavior of those anchor words is inherited from the table’s own birth, not discovered by us. The first comparison, the one that shuffles the matches, defends against the cheap version of this; it does not remove it. It is a structural limit of any shared-sign test, and it caps what a yes could ever have meant here.

The run finished at 2:40 this morning. The overall measure against the first comparison, the one that shuffles the matches: p = 0.028, meaning a result like this turns up by pure chance about three times in a hundred. Against the second, the one that shuffles sign order: p = 0.0001, once in ten thousand. By the rule fixed in advance: yes, the credit belonged to the table itself. The first positive signal on the sounds this project has produced. Not one of the fourteen features taken alone beat both comparisons: four held up against the order shuffle, none against the shuffle of the matches. We noted that and moved on to the review.

The review is the other rule of this phase, also fixed in advance: no result gets interpreted, by us or by anyone, before it passes reviewers who did not produce it. First two reviewers starting from zero, without our context, then an outside AI from a different family, the same protocol that felled a wrong number of ours in entry #007.

Here is what the review found. The source disagrees with itself. Younger’s page carries a caption: “secure syllabograms: DA, I, JA, KI, PA, PI, RO, RI, SU, TA, O”. Eleven values. Our mechanical rule had counted fifteen, because the grid holds four more bare cells (RA, TA2, SWA, TWE) than the caption’s list. The rule existed to keep our judgment out of the table, and it worked, and it still smuggled a judgment in: the choice of which part of the source to believe. We reran everything on the source’s own eleven signs. The chance probability went from 0.028, three in a hundred, to 0.102, one in ten: above the bar. The verdict flipped to no.

Two more deflations, both checked by rerun. Neither fact will surprise anyone who has read these tablets; both correct our own headline numbers. Of the 233 usable words, 185 are one sign long, and a one-sign word contributes exactly nothing to a measure that compares initial against final position. The real evidence base is 48 words, 28 distinct words. And a single word, KI-RO, carries three quarters of the whole effect. The result survives its removal (the chance probability stays at 0.025, recomputed on the same words), so “it is all one word” would be the wrong headline. But 48 words is the honest base, and we published 233.

The outside judge caught one more thing, and it was aimed at us, not at the data. Our internal reviewers had proposed to fix the discrepancy with a new test, declared in advance, on the eleven-sign set. You cannot declare a test in advance when you already know its answer, and we already knew the 0.102. So the rerun is filed as what it is, a check done after the fact, with that label in the first line of the report.

The final verdict for Phase 3, round one: inconclusive. Not a discovery. Not a clean no either; the overall result beats both comparisons on one defensible reading of the source and falls below the bar on the other. No feature gets a name, no sign gets a reading, and nothing here says a word about what language is underneath. That bar has not moved.

Four bare cells in someone else’s grid were enough to flip our first yes. Nine hours from the run to the reversal, most of it spent by our own review taking the result apart. That is what the review is for.

Entry #007 published a shortlist of thirty-two commodity-word pairs this morning. Nine are left. And not one measured number changed.

We commissioned an outside model from a different family, gave it none of our context, and told it to break Phase 2. The attack that landed was aimed at #007, and it graded fatal. The charge: we chose the family, the list of questions to count, after seeing the data, then raised the bar for its size as if we had not.

The machinery is worth spelling out, because the failure hides inside it. Test many hypotheses at once and a few will look good by luck, so you correct: divide the bar by how many questions you asked. The correction is only as honest as that count. Ours counted forty pairs. But forty was not how many questions we asked. Forty was how many places we had already seen something, because a pair only entered the list once it had co-occurred three times or more. We let the answer pick the question.

Fixed before looking, the family is every pairing of a commodity-sign and a word-type that are each frequent enough to test, whether or not they ever appear together. That is 1,728 pairs. 1,504 of them never co-occur at all. Forty-three times what we published, and the bar divides by all of it.

Nine pairs clear. Twenty-three of the thirty-two do not. Nothing new arrives.

The part worth sitting with: the p-values, the measures of how unlikely each pair is by pure chance, are identical. All forty of them, to the digit, unchanged from the page we published. No measurement moved. The divisor moved. Counting your hypotheses only where you already found something inflates your discoveries, and it does it quietly, with the same data and the same arithmetic and no step anyone could point at and call a lie. That is the whole failure, and it went through a ledger we froze in advance, in public, with the post-hoc motivation written into it in plain words. We caught the thing we were watching for and missed the thing underneath it.

One more, and it does not flatter us. The ledger described that family as the pairs co-occurring above zero. The real filter was three or more. The circularity was worse than our own record of it.

What stands: entry #006’s strict verdict, zero pairs surviving the strictest correction, was a separate test and is untouched. What does not stand is the number in #007, struck there now and pointing here. The flawed test, code P2-16 in our public test ledger, is retired with the reason attached, never deleted, and the corrected version, P2-20, replaces it.

A limit on the corrected test too, in the same spirit. With 1,728 pairs the bar at the top three spots of the list falls below the smallest value ten thousand reshuffles can measure. It does not bite here; the nine survivors sit well past that point of the list. But from here on the reshuffle count, not only the corpus, sets how far down we can reach.

While we were counting sign frequencies, the news kept arriving from the other direction. In June a write-up reported that Tom Di Mino, an AI engineer, had established that Linear A is an extinct Semitic language, a precursor to Biblical Hebrew, Arabic and Aramaic. It credits him with a lexicon of 408 words and readings for 40 signs, thirteen of them previously unknown, and says linguists at Rutgers and Cambridge are looking at the claims. Almost none of that is in his own words. The piece carries no byline and quotes him nowhere, and the one claim he makes himself, on his GitHub profile, is shorter and stronger: officially deciphered Linear A. A nine-page draft exists and has not been submitted anywhere. It is a serious person making a big claim with the same tools we use: Python, a pinned corpus, an AI assistant. We cannot referee his work, because none of it is published. No table, no code, no preprint. What we can do is measure the bar such a claim has to clear, on the same corpus, in public.

Here is the experiment. Take the 332 words of the corpus that can be written out in full in sounds, reading Linear A signs with the values of Linear B, the deciphered sister script: the same borrowed hypothesis behind every decipherment built on word lookalikes. Now invent a language: 400 roots of two or three consonants, drawn from a hat. Match them against the corpus the way this genre of attempts does, with flexible sound rules that let related consonants stand in for each other and let a root hide inside a longer word. Count the hits.

A lexicon of pure noise translates 80.6% of Linear A. We ran it a hundred times with a hundred different fake lexicons; the worst of them still covered 77%. Drop the flexible sound rules and demand the exact consonants: 62%. Demand that the whole consonant skeleton equal the root exactly, no skipped letters, no substitutions: one word in five still finds a meaning in gibberish. A 408-word lexicon with hundreds of internally consistent matches is not evidence of Semitic, or of anything. On a corpus this small, with rules this generous, it is the expected result of nothing at all. Cyrus Gordon made the Semitic case on this kind of fuel in 1957, kept making it for decades, and the field never accepted it.

None of that proves Di Mino wrong. Semitic Linear A is a legitimate hypothesis with a long pedigree, and the mainstream position, that Minoan is a language isolate, is a default, not a theorem. The problem is narrower and fixable: the pipeline is upside down. Announce first, publish never, verify later is the exact reverse of what the recurring libation formula makes possible. That formula appears at five different sanctuary sites. Derive your readings from four, then read the fifth in front of witnesses. His claim’s own structure invites the test, and only he can run it, because only he has the lexicon.

We hold ourselves to the inverted version. Entry #007 shows what that looks like in the small: a correction we chose after seeing a result is labeled as such in a public ledger, frozen before it ran, and its output is called a shortlist, not a discovery. Two other projects live by the same order of operations: an independent twin analysis, and a falsification harness built specifically to grade decipherment claims. We have not cross-checked against either yet. The tools are not the divide; we lean on an AI assistant as much as anyone. The divide is whether the check comes before the applause. When his table becomes public, ours is the kind of harness it should be run through, and we will say here what comes out, whichever way it goes.

Entry #006 ended on a non-result, reported because it failed. Thirty-two of forty commodity-word pairings beat chance taken one at a time, and none survived our strictest bar, the one you raise when you ask many questions at once. We could have quietly tried a friendlier bar next and reported whichever looked better. That move has a name in the scientific literature, and it is how a century of decipherments went wrong.

So we did it in the open instead. Before running anything this morning we declared a second test in the ledger, the public list where tests are put in writing before they run: same data, same ten thousand reshuffles, same measures, only a different bar, one that accepts a small expected share of false positives rather than shutting them out almost entirely. The ledger entry says in plain words that this second test was motivated by the first result, after seeing it. It is a weaker net and it was cast second. Anyone can read that before reading the outcome.

The outcome: all thirty-two pairs come back. The cut is not even close to the threshold, which means the choice of threshold is not doing the work. The strongest excluded pair sits at nearly twice the p-value of the weakest included one. Under the strict rule of #006 these pairs are still not discoveries. Under the weaker, declared rule they are candidates: a shortlist to retest on data we have not touched, in Phase 3.

This paragraph is retracted. Measured against a list of only forty pairs, chosen after seeing the data, it counted thirty-two. Against the honest list declared in advance, 1,728 possible pairs, nine survive. See the correction in #009.

A separate question, same discipline. Four Linear A signs moonlight: each is both a word-sign and a commodity-sign, the way a fig ideogram and the syllable ni share one shape. The corpus convention routes each occurrence one way or the other, and convention is not evidence. We walked the raw token stream, in original order, and checked what actually follows each isolated occurrence. Fifty-five cases. In forty-two, a number comes right after: a commodity with a quantity, which is what an accounting ledger does all day. Thirteen have no number next door, and those are the genuinely doubtful ones.

For the doubtful thirteen we asked a witness that does not share our sources. The SigLA project traces each inscription sign by sign from the published facsimiles and tags every occurrence with a role. Seven of our thirteen are in their pages, and all seven are tagged as commodity-signs there too. Zero read as syllables. The other six are not in SigLA at all, so for those the honest record is three words long: we don’t know. One bonus find: in tablet HT 88 our source and SigLA read two different signs in the same slot, and both call it a commodity-sign. That disagreement goes on the pile for epigraphic review, next to the two the QA pass found yesterday.

Phase 2 is now closed. Frequencies, positions, collocations, null models, an exploratory shortlist, and a routing audit, every number regenerable from one script each. Phase 3 is where readings start, which means Phase 3 is where fooling ourselves gets easy. The ledger comes with us.

The site has a house rule: every result gets tested against chance, and if it also shows up in shuffled data it is not a result. Entry #005 flagged a pattern and left it as a promise. This is the promise coming due. AB008 opened words 160 times and closed them 15; AB076 did almost the exact reverse. The question is whether a corpus with no structure at all could throw up splits that lopsided by luck.

Same discipline as last time. Four new tests went into the ledger, the public list where we declare measurements before making them, saved before even the code that runs them. The test keeps every word intact and shuffles only the order of signs inside it, ten thousand times over. Each word keeps its own signs, its length, its damaged spots; all that moves is the order. If AB008 sits at the front by accident, a random reshuffle should pull it off the front about as often as not.

It does not. Not one of the ten thousand shuffled corpora reproduced AB008’s front-loading, and eleven other signs cleared the same bar even after we raised it for having checked all 291 signs at once. AB076’s gentler pull to the back turned up by chance in twenty-three reshuffles out of ten thousand. Position in the word is carrying information. What kind of information is a Phase 3 question. That it is not noise is settled now.

A second test, same logic on different data: do commodity signs and numbers travel together more than chance allows? They do, and not by a little. The four most common commodity signs show up next to a number far more often than the one-in-four base rate, the top ones almost every time they appear at all. That is what an accounting archive looks like from the inside. Where a good is recorded, a quantity follows.

And one result that failed its own test, reported here because it failed. Specific pairings of a commodity sign with a particular word looked strong taken one at a time: thirty-two of forty beat chance on their own. Under the bar raised, as it must be, for having sifted through thousands of possible pairs, none survived. A handful of single-sign words are so common they attach to everything, and they set a bar the real pairs cannot clear. By the house rule that is a non-result, not a discovery. Every number, the passing tests and this one, is in the run’s report.

Phase 2 opened with a false alarm. Our scouting notes said the source of our data had fixed eight transcription errors after our copy, so the plan was to re-download everything and redo the audit. We checked first. The fixes were already inside our copy: the note had compared the dates of the fixes without checking the hour of our import, which happened less than an hour after the last fix landed. Two commands of verification saved a pointless re-import, and the record is corrected where the claim used to be.

Then we locked the method. Before computing a single statistic we wrote a public ledger of tests: eleven measurements declared in advance, each with its counting unit, its policy for damaged signs, its thresholds, saved before looking at any result. A ledger cannot stop us from fooling ourselves. It does make the fooling visible in the edit history.

The distribution is brutally top-heavy. Twelve sign types out of 291 carry 41.6% of the readable text, while 183 types appear fewer than five times and 112 appear exactly once. Position matters too: AB008 starts words 160 times and ends them 15, while AB076 does the reverse (2 starts, 38 endings). That kind of asymmetry is what a language with real morphology tends to leave behind. In Phase 3 it becomes something to test, not something to assume.

One honesty note, since these counts wear the Fact badge: a fact here means “reproducible by anyone, with one script, on the exact version of the data we locked (code 083cb0c)”, not “certain truth about a 3,500-year-old tablet”. Two single-sign readings are still disputed between our source and SigLA, and upstream can correct more errors at any time. If the data moves, these numbers move, and this log will say so.

The QA pass came back. We pulled 25 inscriptions across 12 excavation sites and every type of written object, and checked them sign by sign against SigLA, an independent digital catalog that redraws each inscription from the published editions. 17 matched exactly. Six of the eight differences turned out to be philosophy, not error: our corpus marks a gap where SigLA risks a guess, or the two catalogs name a ligature differently.

That leaves 2 genuine disagreements, each about a single sign, out of 7,742 in the corpus. Both are recorded with the exact facsimile page a specialist would need to settle them. Verdict: the statistics phase can proceed.

Domain bought, DNS pointed, this page shipped. Meanwhile a quality check is running tonight: a sample of our inscriptions, spread across sites and object types, checked sign by sign against SigLA, an independent catalog that redraws each inscription from the published editions. Results in the next entry, including the misses.

Imported all 1,721 inscriptions from lineara.xyz, locked to an exact version of the source, with file fingerprints recorded so every count can be redone. Facts, hypotheses and the source’s own conjectures now live in three separate files, so speculation can never leak silently into the data.

Also: our own scouting notes claimed the source had 1,683 entries. Two independent downloads with identical checksums say 1,721. The recon miscounted, so we corrected the record instead of explaining it away. Expect more entries like this one.

Before writing code we tried to kill the project’s core assumption. One casualty: the database we planned to build on turned out to store its data in a format we could not use, so we switched to lineara.xyz and kept the other as an independent cross-check. Cheaper to find out in week one than in month six.