# Where else could a BIP-39 word live? A survey of every non-essence word source Scope: the 512 artworks are GAN 1 through GAN 512. GAN 0, the Pandimensional skullGAN, is a SuperRare 1/1 released on a different contract and is excluded from every count below. All figures come from `master_dataset.json` (513 rows), `all_onchain.json` (296 tokens whose on-chain attributes were actually fetched), the four Arweave zips in `arweave_dl/`, and the 109 pages of the site under `site/txt/`. ## Why this question is worth asking, and where the analogy breaks The Act 1 clue lives at `site/txt/news__cryptoartgans.txt`, posted February 5 alongside the four new cryptoArtGANs (Colborn, SuperRareRoses, Brinkman and StellaBelle, which are GAN 445 through 448). The clue reads, in full: "The seed phrase is 11 glitches ending with a random word." The prize is the Magic Internet moneyGAN, GAN 445, sitting in the wallet. The reason to distrust the essence field is the Winter Solstice puzzle, where Pindar's third clue said "Their essence is just a distraction, they hide their true BIP39 nature elsewhere," and the published solution at `site/txt/news__pattern-revealed.txt` confirms it: every one of the twelve words came from the **Date** attribute, and the order came from mint order. Two of the twelve were plain-sight attribute values ("Day" and "Night"), four were decoded from dates that fell on an equinox or a solstice, and six were month names that happen to be BIP-39 words. That precedent tells us exactly what a viable non-essence source has to look like. It has to be one attribute; that attribute has to carry many different values; and enough of those values have to be BIP-39 words, or decode to BIP-39 words, that eleven or twelve chosen artworks can each contribute a **different** word. The Solstice answer used twelve distinct words, and so did the earlier robot-drop puzzle solved by noohp. Every published bitGAN solution has used distinct words. The finding that dominates this whole report is that on the Act 1 side the attribute schema is too poor to support that. The bitGANs expose exactly ten on-chain attributes — GANg, Type, SubType, Essence, Eyes, Dimensions, Drop, Date, HiddenType and Name — and of those, only Essence has both broad coverage and enough distinct values. The Solstice drop had a rich Date attribute to hide in. The Act 1 pieces do not have an equivalent. ## Source-by-source coverage ### 1. Essence (the baseline every existing set uses) 511 of the 512 artworks have an essence that is literally a BIP-39 word. The single exception is GAN 191, whose essence is "err0r" with a zero for the letter o, and which decodes to the BIP-39 word "error" under leetspeak. Essence therefore covers all 512 and offers roughly 490 distinct words. It is the only field in the collection with both properties. Any selection rule at all can be worded from it, which is precisely why it cannot by itself discriminate between competing theories. Worth recording for the other workers: essences run in near-perfect alphabetical order inside the earlier drops and break down in the later ones. Adjacent ascending pairs run 125 of 127 in the skullGAN drop, 15 of 15 in the October spookyGAN drop, and 89 of 91 in the Solstice drop, but only 37 of 63 in the robot drop, 85 of 143 in Ghosts in the Machine, and 36 of 67 in Act 1 Review. ### 2. Title, and the title with the GANg suffix stripped Taking the Title column exactly as written and splitting it on non-alphanumerics, 139 of 512 titles contain a BIP-39 word, and 127 contain exactly one. Once you also strip the "GAN" suffix from the GANg token, which is what the `TitleWords_all` column does, coverage rises to 356 of 512 and the distinct-word ceiling rises to 47. The coverage number flatters the source badly. Of those 47 distinct words, "skull" accounts for 159 artworks and "ghost" for 81, and after those two the list falls away fast: test 38, slender 31, monster 28, frozen 24, vivid 19, space 15. Almost every title word is simply the piece's own GANg or SubType read back to you, which is the opposite of hiding a word elsewhere. The part of this source that is genuinely interesting is the tail. Thirty-one artworks carry a BIP-39 title word that appears in **no other title in the collection**: artist (GAN 2), whale (5), spirit (13), heart (34), ancient (44), warrior (55), awkward (141), cat (147), dog (159), evil and olive (174), wolf (180), pumpkin (198), angry (204), solar (209), hungry (259), acid (264), bird (268), method (282), lunar (300), dry and ice (327), eye and swarm (370), spin (379), fire (401), found (404), twenty (420), magic (445), rose (446), bean (449), vibrant (494), cloud (496) and quantum (512). These thirty-one split across the six drops as 6, 5, 2, 6, 6 and 6. **Verdict: this source could carry the puzzle**, but only through its thirty-one-artwork tail, and only if some rule picks eleven of those thirty-one. No trait value, no glitch flag, and no eleven-long run in mint order picks exactly eleven of them. ### 3. Name Thirty-eight artworks carry a Name in the sheet and thirty carry one on chain among the 296 tokens that were fetched. Sixteen of the thirty-eight sheet names contain a BIP-39 word, and fifteen of the thirty on-chain names do. The distinct-word ceiling is 19. The named artworks fall into three families. The nine nagaGAN palindromes from the September robot drop are GAN 129 RoboboR, GAN 130 qubitGAN, GAN 141 Dr Awkward, GAN 147 Taco Cat, GAN 159 Dog God, GAN 162 Star Rats, GAN 174 Evil Olive, GAN 180 Flow Wolf and GAN 192 CalcuclaC; five of the nine yield a word (awkward, cat, dog, evil or olive, and wolf) and four yield nothing. The Ghosts in the Machine drop contributes Frosty Ghost, Dry iceGAN, Eye Swarm, Twinvader, Test Test, Space-Clops, 404 Not Found, Mint Bot, Four twentyGAN and Phantom. The Act 1 Review drop contributes the four cryptoArtGANs plus Bean, Testy McTestGAN, Horned skullGAN, Cloud Skull, Dark Rainbow and Quantum Glitch. Coverage of sixteen is too thin to word an arbitrary eleven-artwork selection, so Name cannot be the source for a rule chosen on other grounds. It is strong in the other direction. **Name is the only non-essence source in the collection that lands on exactly eleven artworks with eleven distinct words under a motivated filter.** If you take the sixteen named artworks that yield a word and remove the five whose word is simply one of the collection's recurring trait words — The Artist giving "artist", Frosty Ghost giving "ghost", Test Test giving "test", Space-Clops giving "space" and Cloud Skull giving "cloud" and "skull" — what remains is GAN 141, 147, 159, 174, 180, 327, 370, 404, 445, 449 and 512, carrying awkward, cat, dog, evil, wolf, dry, eye, found, magic, bean and quantum. Two independent motivations land on that same eleven. It is also the intersection of the thirty-one globally-unique-title-word artworks with the sixteen name-word artworks, minus the two whose word is generic. The same filter applied to the on-chain Name attribute instead of the sheet column gives a slightly different eleven, because the on-chain data adds Bird monsterGAN and Fire Squid and drops Quantum Glitch and Dry iceGAN. Note the calibration point honestly: landing on eleven proves nothing on its own. What makes this one worth trying is that the filter was not tuned to produce eleven — it was "drop the names whose word is just a trait word," and eleven is what fell out. ### 4. Trait values: GANg, SubType, Type, Eyes, Dimensions, Level, Drop, HiddenType This is the family the Solstice precedent points at, and it is the family that fails hardest. **Type** is a BIP-39 word for 252 of 512 artworks, which sounds like usable coverage, but there are only eight distinct Type words in the entire collection: ghost (105 artworks), skull (75), monster (27), cloud (18), smoke (17), artist (7), alien (2) and mechanic (1). Eight distinct words cannot produce eleven distinct seed words. **SubType** is a BIP-39 word for 121 of 512, again across only eight distinct words: slender (36), frozen (31), vivid (26), silver (11), twin (10), shy (4), quantum (2) and skull (1). The same ceiling applies. This is the same observation the Solstice discord made — theory 11 in `news__pattern-search.txt` counted thirteen BIP-39 subtypes across the Solstice drop specifically — but across the full 512 the count is eight. **GANg** yields nothing as written, because every value ends in "GAN". Strip that suffix and you get skull, ghost, test, monster, space, whale, spirit, slender and ice as BIP-39 words, which is nine distinct words and is already counted inside the title source. **Eyes, Dimensions, Level and Drop** produce no BIP-39 word at any artwork. Eyes and Dimensions are integers, Level is "1" or "infinite", and Drop is a date string or a drop name. **HiddenType** takes twelve values, none of them a BIP-39 word. **Date** is present on only 69 artworks, all of them in the Solstice drop, and yields sixteen artworks with a word across eight distinct words. It is the attribute that solved the Solstice puzzle and it is exhausted; it does not reach the Act 1 pieces at all. **Verdict: no trait value field can carry this puzzle** if the answer uses eleven distinct words. BIP-39 does permit repeated words, so for completeness the candidate file includes the two trait groups that hit exactly eleven — the eleven Silver-subtype artworks, all giving "silver", and the eleven Crypto-subtype artworks worded from their Type attribute, giving seven "artist", three "ghost" and one "cloud" — but I would rank both near the bottom. ### 5. Drop descriptions and drop themes `M_DropTheme`, `M_DropPuzzle`, `M_OriginalDrop` and `M_Notes` are prose fields, and the first two contain BIP-39 words for all 512 artworks precisely because they are prose. They are useless as a per-artwork source because every member of a drop shares the identical string, so the field cannot distinguish one artwork from another. Not a single artwork has exactly one BIP-39 word in `M_DropTheme`. `M_OriginalDrop` has one curiosity: eleven artworks carry the string "Ghosts in the Machine" and the only BIP-39 word in it is "machine". Those eleven are GAN 490 through 496 and 501 through 504. Eleven identical words, so the same objection as the trait groups applies, and the eleven is an artefact of which Act 1 Review pieces were reprises rather than a designed count. The drop write-ups on the site are narrative, not per-artwork. The one page that looks like a per-item word list, `site/txt/journal__page010.txt`, is the podGAN rarity CSV, and podGANs are a separate collection of 1024 with their own "seed" attribute that genuinely does carry a word per item. That is a real precedent that Pindar mints word-valued traits, but those words belong to the podGANs and not to the 512 bitGANs. ### 6. Filenames and Arweave codes `ArweaveFilename` is populated for 339 of the 512. Among those, 144 contain a BIP-39 word, with a distinct-word ceiling of 19. The Solstice-drop filenames are dominated by seasonal boilerplate: summer appears in 15 filenames, autumn in 12, winter in 11, spring in 11 and update in 9. Those are the same four season words the Solstice puzzle used, but they are visible here in the filename rather than decoded from the date, so they are a restatement of a solved puzzle rather than a new source. The interesting part is the first drop. The 127 skullGAN filenames use hand-written descriptors that never appear in the on-chain metadata: ancient, charred, scholar, whale, smooth, blob, spirit, elder, beanie, heart, warrior, shark, triplet. Setting aside the two generic ones, skull and ghost, exactly eleven of the BIP-39 words in that drop's filenames remain — ancient, elder, heart, silver, slender, smoke, smooth, spirit, twin, warrior and whale. Five of those eleven sit on a single artwork each: whale on GAN 5, spirit on 13, elder on 16, heart on 34, warrior on 55. The other six have to be assigned to an artwork by choice, which weakens the set badly, but the vocabulary itself is a genuine "hidden elsewhere" channel, since "elder" on GAN 16 and "smooth" on GAN 89 appear nowhere in those pieces' titles, traits or essences. Seventy-two artworks have a filename BIP-39 word that appears nowhere in their own metadata. `M_ArweaveCode` and `CLUSTER` hold the same twenty-five internal batch labels — roses, brinkman, stella, colborn, dgans, highdef, skullall, aa, bs, line, lines, test and numbered clusters. Only "test" is a BIP-39 word, on seven artworks. `M_Hexadecimal` holds only the eighteen octal and hex renderings of the spookyGAN numbering ("0o0" through "0o17", "0x20", "0x21"), which contain no letters to decode. `M_BinHexConv` holds the decimal value of the binary-titled robot pieces, running 0 to 75, which are numbers and not words. **Verdict: filenames are a real and under-explored channel, but coverage is a floor rather than a ceiling**, because 173 artworks have no recorded filename at all. ### 7. Derived words: reversal, leetspeak, hex and binary All four of these turn out to be nearly empty, which is worth stating flatly because they are the most attractive theories on paper. **Reversal** produces a BIP-39 word at exactly two artworks in the entire collection: GAN 159 "Dog God", where "god" reverses to "dog", and GAN 180 "Flow Wolf", where "flow" reverses to "wolf". Both are palindromic nagaGANs and both already yield the same word by simply reading the title forward. The famous reversal, GAN 48 whose on-chain SubType is "wobniaR", reverses to "Rainbow", and "rainbow" is **not** a BIP-39 word, though it is a SLIP-39 word. **Leetspeak** produces a BIP-39 word at five artworks, and only one of them is real. GAN 191's "err0r" decodes to "error". The other four are my decoder reading digit strings inside titles and filenames — "136" and "139" as "leg", "463" and "493" as "age" — which are numbering artefacts and not words. **Numeric indexing** does not encode anything. Looking up the BIP-39 word list at the artwork's own number reproduces its essence at exactly one artwork out of 512, GAN 360, which is what chance predicts. Looking up at the mint-order index reproduces the essence at zero artworks. Looking up at the decimal value of the robot drop's binary title reproduces the essence at zero artworks. **SLIP-39 substitution** adds almost nothing. Across every text field in the collection, the only values that are SLIP-39 words but not BIP-39 words are "rainbow" (the SubType of 82 artworks and a token in 37 filenames), "scholar" (one filename) and "phantom" (one name). ### 8. GIF file internals The byteGAN puzzle was solved by pulling Photoshop layer names out of a GIF's embedded metadata, so this is the most obvious place for the words to be hiding. I walked the GIF block structure of all 404 files in the four Arweave zips. Every one of them carries exactly two extension blocks beyond the frame-control blocks: a NETSCAPE2.0 looping block, and an XMP packet. The XMP packets are pure Adobe boilerplate. Each holds a DocumentID, an InstanceID, a DerivedFrom pair and a CreatorTool string, and nothing else. There are no layer names, no titles, no descriptions and no keywords. The CreatorTool values are clean and fall into five buckets: Photoshop 2022 Windows on 180 files, Photoshop 2021 Windows on 165, 23.1 on 29, 22.3 on 26 and 22.5 on 4. The DerivedFrom source IDs are almost entirely unique per file: 343 files have their own, 59 have no DerivedFrom at all, and exactly one pair of files (GAN 14 and GAN 75) shares a source document. One warning for anyone who repeats this. XMP inside a GIF is written with a magic trailer that lets naive sub-block parsers walk it, and a parser that treats it as ordinary GIF sub-blocks silently drops one byte in every 256. That makes the CreatorTool strings look corrupted — "Adob Photoshop", "Windws", "Windos", "www.w.org" — and I briefly believed I had found a new glitch family. The corruption is entirely in the parser. Read the packet raw, from after the "XMP DataXMP" marker to the closing "