# Lens 4: puzzle design and cipher history Swarm of 21 September 2026. Ideas only. I derived no wallet, ran no sweep harness, touched no network and edited no existing file. The only arithmetic I ran was reading the data files, looking up sorted word sets in `theories/_registry.json`, and the BIP-39 checksum (one SHA-256 on the phrase's own bits, the same "checksum arithmetic only" used in D75). ## The starting point: how Pindar built his earlier puzzles I read his own write-up of the podGALLERY puzzle (`site/txt/news__podgallery-solution.txt`, 3 December 2021, two months before the Act 1 clue). His four stages were: 1. Find the markers. "There are 14 duplicates between the 6 pods - 12 that matter, and 2 meant to be a distraction." 2. See the pattern. The markers sat "in the First and Last square of each pod", which gives two pieces from each of six pods, twelve in all. He adds that the first and last pieces "were the only ones out of sequential order on some of the pods", and "It didn't matter how you discovered the pattern." 3. Read the words off the twelve pieces' attributes, in order. 4. A twist on the words (they came from the wrong wordlist). The Solstice puzzle used the same end-pinning: "It begins with sunrise and it ends with sunset" (#209 and #300, the first and last pieces of that drop). So his habit, stated twice in his own words, is: **markers that repeat or break the count, sitting at the first and last place of each container, twelve pieces from six containers.** Act 1 has six drops. That is the fact my two best ideas rest on. A design note that matters for every idea: the maker of a seed puzzle does not need eleven chosen words. Any eleven words work, because he can always finish the phrase with one of the 128 twelfth words that make the checksum pass. That is what "ending with a random word" describes. So the words can be whatever falls out of the pieces he marked. The skull essences prove it: skull #n simply carries wordlist word n. --- ## Idea 1 (best). The first and last piece of each of the six drops **Side:** all DATA (the counters printed in the names, and the drop boundaries). - **(a) Selection.** "The glitches sit where my markers always sit: in the first and last place of each drop." Six drops give twelve pieces: #1 and #128, #129 and #192, #193 and #208, #209 and #300, #301 and #444, #445 and #512. - **(b) Word.** Each piece gives its essence, read as printed: abandon, avocado, alpha, zero, glue, grape, avoid, brief, exact, danger, collect, final. - **(c) Order.** Piece-number order. For these twelve pieces mint order is the same order (I checked the mint slots: 66, 203, 208, 280, 288, 305, 309, 411, 413, 582, 660, 662), so there is only one natural order. - **(d) Shape.** Eleven words plus a twelfth. The twelfth is either "final" (#512, the twelfth boundary piece) or any of the 128 legal twelfth words. One ordinary harness run over the 128 legal twelfth words covers both, because "final" is one of the 128 (see below). **What in the data teaches it.** - podGALLERY: six pods, first and last square of each, twelve pieces. Solstice: first and last piece pinned by the hint. - In three drops the first and last piece are visibly built as a pair. Robo: #129 RoboboR and #192 CalcuclaC are both palindrome names, with essences "alpha" and "zero". Spooky: #193 and #208 share the name Ghostly spookyGAN, and their essences "glue" and "grape" are the first and last words of that drop's clean wordlist run (800 to 815). Solstice: #209 Solar and #300 Lunar, the collector's sure pair P03. - Four of the twelve boundary pieces have an unusual frame count: #129 (108 frames), #192 (44), #445 (88), #512 (63). Only 29 of the 512 pieces have one. The chance of four or more landing in a random twelve is about 3 in 1,000. These are both ends of the robo drop and both ends of the Review drop. - Eleven of the twelve carry at least one of the site's glitch theme tags (251 of 512 pieces do; the chance of eleven or more in a random twelve is about 2 in 1,000). Examples: #128 is a sheet-versus-chain error piece with a filename typo; #129 holds the displaced word "alpha"; #193 wears the hidden-type letter F that belongs only to robots; #208 has the doubled space in its chain name; #445 was minted out of order. The one piece with no tag is #444 "danger", and its oddity is that "danger" is wordlist word 444, its own piece number. **What I could check without a wallet test.** - The registry has none of the twelve possible eleven-word subsets of these twelve words, and none of the 78 subsets when "universe" is added. As far as the registry shows, this set has never been run. It looks too simple to have been missed, so please double-check that with your own lookup. - **The full twelve-word line, in piece order, passes the BIP-39 checksum:** `abandon avocado alpha zero glue grape avoid brief exact danger collect final`. "final" is one of the 128 legal twelfth words after those eleven. A random twelfth word passes one time in sixteen. Honest count of my forking paths: I ran this checksum on thirteen fixed twelve-word lines today (listed at the end), and this is the only one that passed. One pass in thirteen tries is what chance would give a little more than half the time, so the pass is not proof. It is still worth noting that the one line that passed is the one with the clearest design reason, and that "final" is the single most fitting last word in the collection. The variants I checked that do NOT pass: the same twelve reversed; firsts then lasts; Solstice words "day" and "night" swapped in for #209 and #300; and drop boundaries taken by mint slot instead of piece number (which would swap in #359 "bright" and #447 "cloud"). **Phrases a test needs.** One ordering × the 128 legal twelfth words. If you want margin: also the reverse, firsts-then-lasts, and each eleven of the twelve in piece order, about fifteen orderings in all. **Why Pindar would do it this way.** It is his own published grammar, used twice, moved from six pods to six drops. It needs no judge. It spans all 512 pieces, as he told the collector. It could only be solved "after some more drops", because the last drop did not exist yet. And a designer who controls #445's and #512's essences can nudge them until a fitting word such as "final" is checksum-legal. **Weakest point.** The clue says "11 glitches ending with a random word", and "first and last of each drop" is not what anyone would call a glitch. #1 and #444 in particular carry no planted error we know of. Also "final" as the twelfth word is fitting, not random. The checksum pass has a 1-in-16 chance rate and I looked at thirteen lines. --- ## Idea 2. The twelve Review pieces that count in binary ("14 markers: 12 that matter, 2 a distraction") **Side:** all DATA (the counter printed in the title). - **(a) Selection.** "In the Review drop every piece counts n/512, except fourteen that count in the wrong system. Twelve of them matter and two are a distraction." The twelve are the binary-counted pieces #457 to #464 (1000000 to 1000111) and #497 to #500 (1001000 to 1001011). The two distractions are the hex-counted #465 (0x20) and #466 (0x21). - **(b) Word.** Each piece gives its essence: poem, soup, small, tonight, buyer, never, sorry, torch, define, safe, sound, escape. - **(c) Order.** The counter's own order, 64 up to 75. This is also piece order and mint order. - **(d) Shape.** Eleven words plus a random twelfth. The natural eleven is the first eleven counters (drop #500 "escape"); the safe test is each eleven of the twelve. **What in the data teaches it.** podGALLERY's count matches exactly: fourteen markers, twelve that matter, two distractions, and markers that can be found as "the only ones out of sequential order". In Act 1 Review the fourteen oddly counted pieces are the only ones out of the n/512 count (`primary_facts.ODD_NUMBERING`), and they split twelve and two. The twelve binary counters are twelve consecutive values, which is the length of a seed phrase. Four of the twelve (#462, #463, #464, #497) are four of the collector's six colour-flash pieces. Two of the three solution-block words that sit out of wordlist order (soup #458, sorry #463) are on binary-counted pieces. **Checked without a wallet test.** None of the twelve eleven-word subsets is in the registry. The pointed-at readings of these counters HAVE been tried: the counter read as a wordlist index (words 64 to 75, "among … announce") is in the registry for all twelve subsets, and the counter with its leading 1 dropped, pointing at robots #129 to #140, is in the registry for four of its twelve subsets. The pieces' own essences were never run as a set. The fixed twelve do not pass the checksum forwards or backwards, which fits "eleven plus a random word" and rules out "all twelve as printed". **Phrases a test needs.** Twelve elevens × counter order and its reverse = 24 orderings × 128. Finishing the robot reading (the eight untested subsets) adds 16. **Why Pindar would do it this way.** Same grammar as Idea 1, and the robo puzzle before it was also about binary-counted pieces in mint order. A counter in the wrong number system is a data glitch a solver can see without any judgement. **Weakest point.** There are twelve pieces and the clue says eleven, so one must be dropped and nothing yet says which. A binary counter on a piece that reprises a robot may be plain housekeeping, not a marker. --- ## Idea 3. Each skull is one bit (Bacon's two-typeface cipher on the first drop) **Side:** VISUAL if the bit is "does the animation glitch"; DATA if the bit is "does the label say Glitch". Pick one side per test; do not mix. - **(a) Selection.** "A bitGAN is a bit. Read the 128 skulls as 128 ones and zeros: one if the skull glitches, zero if it does not." - **(b) Word.** Cut the bits into groups of eleven; each group is a wordlist number. There are no eleven "pieces"; the "11 glitches" are eleven eleven-bit words made of glitches. - **(c) Order.** Skull number order, #1 first. - **(d) Shape.** 128 bits is exactly the amount of data in a twelve-word phrase. The first 121 bits make eleven words. The last seven bits (skulls #122 to #128) plus the four checksum bits make the twelfth word, which looks random because the checksum scrambles it. **What in the data teaches it.** The skull drop has 128 pieces, and 128 bits is exactly a twelve-word phrase. Bacon's cipher hides a message in which letters are in one typeface or the other; here the two "typefaces" are glitch and no glitch. The visible frame glitch on skulls follows the prime rule with exactly two breaks, #14 and #75, the byte-identical pair: neither number is prime, both glitch. **Computed (checksum arithmetic only).** - Visible frame glitch (the 28 skulls tagged `visual:frame_splice`): `acquire celery little bacon calm beef melody dune above eyebrow cactus album`. - Label says Glitch (41 skulls): `hamster best pear baby track rose alley educate dog image amount aim`. - Inverted versions: `worry sustain latin upgrade term typical install room youth push that wide` and `nuclear turkey fuel upon bonus dumb west ride satoshi modify wave whisper`. None of the four first-elevens is in the registry. The earlier bit-register runs (extraction catalog EC32, 44 sets) used 121-bit slices of all 513 pieces starting from #0, which gives different words. The word "bacon" in the first line is an accident of where the primes fall; I note it only because it is funny. **Phrases a test needs.** Four fixed phrases, or four orderings × 128 if the twelfth is left free. **Why Pindar would do it this way.** The name bitGAN, his binary counters, and "even knowing how to solve it, it was hard": turning 128 animations into bits and then into words by hand is slow even when you know the method. **Weakest point.** It uses only the first drop, which contradicts "only solvable after some more drops" and "spans all 512". The words are produced by arithmetic, not read off a piece, which is unlike his solved puzzles. --- ## Idea 4. The piece that turns up in the wrong drop for its name (déjà vu) **Side:** all DATA (names and drops). - **(a) Selection.** "A glitch in the matrix is déjà vu: something you have already seen. Where a name is shared by three or more pieces, the glitch is the one piece that sits in a different drop from all the others." That gives exactly eleven: #256, #261, #364, #411, #455, #461, #467, #473, #478, #491, #497 (the "drop" rows of `same_names/odd_ones_out.json`). - **(b) Word.** The essence: bike, bitter, matrix, key, trade, buyer, give, false (written FALSE), recycle, weapon, define. - **(c) Order.** Piece number, then mint order. - **(d) Shape.** Eleven plus a random twelfth. **What in the data teaches it.** #364's essence is "matrix", and #364 is a first-drop name (Nebulous skullGAN) that reappears once in the Ghost drop, "Ghosts in the Machine". The set holds FALSE (#473, the only Entangled bitGAN outside Solstice), which fits the collector's reading that TRUE means true to its drop and FALSE means not. It also holds "key" (#411). **Checked without a wallet test.** The eleven-word set is not in the registry. The sister eleven from the same file, "the only error piece in an otherwise clean name family" (#17, #33, #44, #48, #58, #62, #88, #97, #128, #170, #173), is also not in the registry, but it contains three pieces the collector ruled accidents (#88, #97, #128), so I do not put it forward. **Phrases a test needs.** About twelve plain orders × 128. **Why Pindar would do it this way.** "They have been looking for glitches all wrong" fits a glitch that is a repeat and not a visual stutter or a typo. **Weakest point.** The count of eleven depends on a technicality. Twenty-five pieces repeat a name from an earlier drop. The rule only reaches eleven by requiring a family of three or more with exactly one outsider, which drops the seven exact cross-drop pairs and drops #454 and #492 because a Ghost-drop skull already shares their names. It also mixes two directions: seven Review pieces that join an old family, and three old pieces (#256, #261, #411) whose family is otherwise Review. The brief's own warning applies: many rules give eleven by chance. --- ## Idea 5 (a LESSON, not a full selection). The glitched piece points at its clean twin **Side:** all DATA. - **(a) Selection.** "When a piece's label or file is wrong, it is standing in for another piece." The cases: #14 plays #75's file; and five error pieces are one half of an exact name pair whose partner is clean (#4 with #36, #66 with #40, #123 with #99, #281 with #244, #315 with #426). - **(b) Word.** Read the clean twin, not the glitched piece: announce (#75), afford (#36), agent (#40), arrange (#99), believe (#244), upper (#426). TRUE and FALSE are the legend: the glitched half is the false one, so read the true one. - **(c) Order.** By the glitched piece's number (#4, #14, #66, #123, #281, #315), or by the twin's. - **(d) Shape.** Cannot be stated: the rule fills six slots. G1itchb0t (#170 error; clean #172 "robot" and #182 "virus") gives one or two more with a free choice, and the Twin Glitch trio (#47, #59, #73) has no clean member. It does not reach nine. **Why it is still worth writing down.** It is the book-cipher habit from podGALLERY (the word on the marked piece is the wrong word; the right word is one hop away), and it is the one reading of the same-name pairs that says which member speaks without a coin toss. The one-of-each-pair runs (hyp72, hyp81) covered the collector's deliberate pairs, not these name pairs. If a later finding supplies three to five more "stand-in" pieces, this rule is ready. **Weakest point.** It does not fill a phrase, and four of its six error pieces (#4, #123, and the not-sure #66, #315) are ones the collector does not rate as planted. --- ## Idea 6 (closed by computation). Arithmetic on a pair's two wordlist numbers I was asked to compute this on the real pairs. For all 38 exact same-name pairs and all 13 collector pairs I took the two essences' wordlist numbers a and b and looked up the words at b − a, a + b and a XOR b (wrapping at 2048, counting from 1 and from 0). - For the six skull-drop pairs the result is only the difference of the piece numbers, because skull #n holds word n. Nothing new can come out. - For the 32 other name pairs the difference words are: congress cigar empower divide jacket forget abuse acid creek age liberty move foil pool burger extra pig jacket bubble garbage absent nominee opinion antenna someone family pink agent license busy cinnamon bird. Sums: gain glass arrange black chief narrow sustain clarify fork comic buffalo fringe normal agree satoshi pear drastic ancient library search parent artefact depth embark climb narrow brave clarify squirrel keen jar maple. XOR: express foot slender say oak hold actor acid farm ask liberty much grace way dish extra run warfare leader pave busy try piece curve spread family swap agent lonely giant jar jealous. - For the collector's pairs the results are of the same kind (for example TRUE and FALSE give notice, desert, road; solution and key give feed, employ, regret). - I see no theme, no run and no instruction in any of the lists. Fourteen of the 32 difference words are held by some piece against about eight expected, but seven of those fourteen are small numbers that fall in the solid block 1 to 224 where every word is held, so that is explained. - Curiosities, none strong: "problem" (#143) minus "hidden" (#168) is exactly 512, which is the word "divide" (#449). "mystery" (#431) and "number" (#470) give 40 by both difference and XOR, the word "agent" (#40). Three collector pairs sit a power of two apart (#9 and #73 by 64 pieces; #14 and #75 by 64 raw mint slots; #363 and #473 by 128 raw mint slots), but raw mint slots have gaps Pindar did not control (by mint rank those two gaps are 61 and 109), so I read this as chance. - I also read each pair's two essences as a two-word message ("exist twice" on the two Ultra ghostGANs #361 and #393, "unlock under", "mystery number", "hidden problem", "key solution"). Two sets of random pairs drawn from the same drops gave lines just as suggestive ("six solve", "reveal broken", "awkward puzzle", "half correct"), so the essence vocabulary is puzzle-flavoured everywhere and this is most likely the eye finding patterns. "exist twice", "brother echo" (#132, #139) and "couple" (#344) are the only ones that actually talk about being a pair. **Verdict:** pair arithmetic gives noise. I would not spend a test on it. It is a total function, so getting eleven legal words out of it would prove nothing (the catalog's own warning under EC26). --- ## Found again and already tested - The eleven Ghost-drop pieces whose essence is the wordlist word at their own piece number (#307 chaos, #322 choice, #348 clock, #368 combine, #372 common, #380 control, #389 correct, #404 crash, #407 crazy, #429 culture, #444 danger). I found it independently; the registry has it as "ghosts-identity11 / fixedpoint11". Tested, no match. It matters to Idea 1 only because #444 is one of the eleven. - The painted numbers as pointers (#165 "81" to #81, #207 "4" to #4) are in D37 and closed as too small. ## Every fixed twelve-word line I put through the checksum today (for an honest count) 1. Twelve binary Review essences, forwards and reversed: fail, fail. 2. Wordlist words 64 to 75 counting from 1, and from 0: fail, fail (each also reversed: fail). 3. Robots #129 to #140: fail, and reversed fail. 4. Drop boundaries by piece number, piece order: **pass**. Reversed: fail. Firsts then lasts: fail. With day and night: fail. 5. Drop boundaries by mint slot: fail. One pass in thirteen lines; chance alone predicts about 0.8 passes. ## Files I read `theories/swarm_2026-09-21/CONTEXT.md`, `theories/DECISIONS.md` (D32, D37 to D43b, D61 to D95), `~/Desktop/bitganspuzzlebook/pages/brief/BRIEF.md` (Parts 1, 4, 5, 6), `site/txt/news__podgallery-solution.txt`, `theories/brainstorm_fodder.md` (B13 to B16), `theories/transforms_and_mechanisms.md` §3b, `theories/extraction_catalog.json`, `primary_facts.py`, `pairs.json`, `lore_catalog_runs/same_names/site_data.json` and `odd_ones_out.json`, `lore_catalog_runs/deep_review_2/data_plain.json`, `theories/anomaly_ledger.json`, `~/Desktop/bitganspuzzlebook/db/tags.sql`, `theories/_registry.json`.