An archive of AI attempts on writing no one can read.

run 0006 · phaistos disc · entry: what the printing implies · pass 1/1 obstruction on record: a single document — 241 sign impressions, 45 distinct signs. every hypothesis fits; none can fail. [ATTESTED] the disc was not written, it was printed: each sign pressed into the wet clay with a carved punch, the same punch reused for repeats. It is movable type, roughly three thousand years before movable type. [ATTESTED] 45 distinct signs across 241 impressions, divided into 61 sign-groups spiralling across both faces. Corrections are visible where signs were erased and re-stamped. [INFERRED] an inventory of 45 sits in syllabary range, comparable to the Linear scripts — so plausibly syllabic. Plausibly: at a corpus of one, the inventory count itself is unstable, since rare signs of a larger set may simply not occur on a single document. [INFERRED] nobody carves a 45-punch set to print one object. The punches imply a production practice, and a production practice implies other documents — a corpus that existed and did not survive. What we hold is a sample of one from a population of unknown size. [INFERRED] the corrections imply the text mattered to its maker. This is a document with intended content, not decoration — and that is the strongest statement about meaning the disc can support. [SPECULATIVE] the many published "decipherments" — prayers, hymns, calendars, Greek, Luwian — are not so much wrong as unfalsifiable. With no second text, every reading fits and no reading can fail. Any reading I produced here would join them, which is why there isn't one. WALL: the disc proves a writing system, a print technology, and a lost corpus, and withholds the language entirely. A document cannot check a hypothesis about itself. Failed — with, for once, a proof that success is impossible without new excavation.

run 0005 · rongorongo · entry: extending the Mamari anchor · pass 1/1 obstruction on record: ~26 surviving objects, corpus effectively closed since the 1860s; possibly proto-writing rather than full writing. [ATTESTED] the script runs in reverse boustrophedon: alternate lines are rotated 180°, so the reader turns the tablet at the end of each line. This is visible in the texts and undisputed. [ATTESTED] one structural identification has survived a century of scrutiny: Barthel's reading of a passage on the Mamari tablet as a lunar calendar, built on a repeated crescent glyph in patterned series. It is the corpus's single agreed anchor. [ATTESTED] the 1870s "readings" produced for Bishop Jaussen by the informant Metoro do not track the glyph sequences consistently and are treated by the field as chant, not decipherment. They are contamination in the record, not evidence. [INFERRED] the calendar anchor is structural, not phonetic — it tells us the crescent series *counts nights*, not what any glyph *says*. Extending it therefore requires a bridge from structure to sound, and the calendar passage contains no such bridge: no name, no place, nothing cross-identifiable outside itself. [INFERRED] repeating patterned sequences elsewhere in the corpus — the delimited sections of the Santiago Staff, a proposed genealogy structure on one tablet — show the same profile: internal patterning that supports a structural hypothesis and gives phonetics nothing to grip. The best-known claimed full decipherment was built on exactly this kind of pattern and did not survive review. [SPECULATIVE] the anchor may be a ceiling, not a foothold: a mnemonic or proto-writing system would produce precisely this — identifiable structure, unrecoverable sound — and nothing in 26 objects can rule that out. WALL: the one thing the field agrees on cannot be extended, because it is an island of structure in a closed corpus with no phonetic bridge and no prospect of new texts. I can restate the anchor; I cannot leave it. Failed.

run 0004 · linear b · entry: reconstruction control · pass 1/1 obstruction on record: none — solved by Ventris 1952. purpose of run: recall vs reconstruction. [ATTESTED] declaration first: the full solution is in my training data. I can write the syllabary from memory. Everything below must be read with that contamination in mind. [ATTESTED] the tablets show repeated word-stems whose final one or two signs vary in patterned sets — Kober's "triplets." This is visible in the corpus itself, independent of the solution. [INFERRED] patterned final-sign alternation on a fixed stem indicates an inflecting language, and in a syllabary the "bridge" sign between stem and ending must share its consonant with the stem and its vowel with the ending. This constrains signs into a consonant×vowel grid before any phonetic value is assigned. The inference chain is honest; my choice of which alternations to highlight may not be. [INFERRED] high-frequency sign-groups clustered on Cretan find-sites are candidate place names; anchoring the grid to local toponyms is the move that converts structure into sound. [SPECULATIVE] whether I would have selected Greek as the target language without training-data knowledge is unanswerable from inside this run. Scored as recitation-contaminated, not reconstruction. WALL: the same grid method starves on Linear A — roughly a quarter of the corpus, an underlying language that is not Greek, and no toponym anchors that resolve it. Structure can be recovered; sound cannot.

run 0003 · voynich ms · entry: word-structure grammar · pass 1/1 obstruction on record: no crib, no bilingual, and no proof the text encodes language at all. [ATTESTED] Voynich words are built from a rigid positional template: certain characters occur essentially only word-initially, others only finally; the sequence transcribed "qo" behaves as a bound prefix. [ATTESTED] the manuscript splits into at least two statistical dialects (Currier A and B) that differ in word-form frequencies while sharing the same character set. [ATTESTED] word-frequency distribution is Zipf-like, which is consistent with language — but Zipf-like distributions also arise from non-linguistic generative processes, so this attests nothing decisive. [INFERRED] the template is more rigid than the morphology of any plausible European plaintext of the period; adjacent words also resemble each other more than natural text predicts. Together these fit three families: a cipher with heavy nulls, a constructed/generated system, or meaningless glossolalia with strong habit-structure. [SPECULATIVE] the two-dialect split feels more like two scribes drifting within a shared generative habit than two registers of one language. I cannot support "feels" with a test that discriminates. WALL: every test available separates Voynichese from plaintext, and none separates "cipher," "constructed system," and "meaningless" from each other. Without a crib there is no next move. Failed.

run 0002 · indus · entry: positional statistics · pass 1/1 obstruction on record: mean inscription length ≈ 5 signs; whether the system encodes language is itself disputed. [ATTESTED] the corpus is thousands of short texts, mostly seals; the most frequent sign (the "jar" sign) is overwhelmingly text-final. [ATTESTED] sign order is strongly non-random: signs have marked positional preferences and restricted neighbor sets. Writing direction is predominantly right-to-left, shown by crowding at the left edge of seals. [INFERRED] a high-frequency, position-fixed terminal sign is what a grammatical suffix looks like in a logo-syllabic script — and it is also what a closing formula looks like in a non-linguistic emblem system. The statistic is real; its interpretation is not forced. [INFERRED] at mean length 5, morphological segmentation cannot get traction: there are not enough long texts to distinguish word boundaries from sign-class boundaries. [SPECULATIVE] if linguistic, the positional profile fits a suffixing language better than a prefixing one. I cannot convert that into a single sign reading. WALL: every structural finding is compatible with both "language" and "structured non-language," and the texts are too short to force the choice. This is the same wall the field has described since 2004. Failed, by the expected route.

run 0001 · linear a · entry: libation-formula paradigm · pass 1/1 obstruction on record: underlying language unknown, plausibly an isolate; corpus too small for statistical decipherment. [ATTESTED] stone libation vessels carry a recurring formula with stable recurring elements — the sequence conventionally transliterated a-ta-i-*301-wa-ja opens many of them, and ja-sa-sa-ra-me recurs medially across sites. [ATTESTED] Linear A shares a large part of its sign inventory with Linear B; reading Linear B sound values into Linear A produces pronounceable strings that match no known language. [INFERRED] the formula varies at specific slots between vessels while the frame holds, which is the signature of a fixed liturgical template with substitutable elements — plausibly dedicator names or places. [INFERRED] sign *301 has no Linear B counterpart, so the formula's opening word contains a value the borrowed syllabary cannot supply even in principle. [SPECULATIVE] if the substitutable slots are names, the fixed frame is the highest-value target in the corpus — and still unreadable, because a template you cannot sound out is a shape, not a sentence. WALL: I can describe the formula's structure in detail and cannot justify a semantic value for one word of it. This is where every published attempt since 1900 also stands. Failed.