# Lens 2 — The glitch is a label that lies about the picture; instruction words as operations

Swarm of 21 September 2026. Ideas only. Nothing here was run against the wallet, and no existing file was changed.
Every fact below was checked against `primary_facts.py`, `tags.sql`, `anomaly_ledger.json`, `data_plain.json`, `DECISIONS.md` and `_registry.json`.

---

## Part A. What I found before proposing anything

### A1. If a label lies, the true word and the false word are usually NOT seed words

I wrote down, for each of the collector's two-word pieces, the word that is untrue and the word that is true, and checked both against the wordlist.

| Piece | Untrue word | True word | Is either a seed word? |
|---|---|---|---|
| #281 | glitchGAN (it does not glitch) | bitGAN | neither "glitch" nor "bit" is on the list |
| #173 | Smoke (blockchain) | Ghost (sheet) | both "smoke" and "ghost" are on the list |
| #66 | Rainbow (sheet) | Flaming (blockchain) | neither; only the cut-down "flame" is |
| #58 | Glitch (sheet) | Rainbow (blockchain Type) | neither |
| #17 | Ghost (sheet title) | Glitch (sheet Type, blockchain title) | "ghost" only |
| #39 | skullGAN (sheet species) | cyclopsGAN | "skull" only |
| #509 | skullGAN or roboGAN (it is half of each) | both half true | "skull", "robot" |

Only #173 gives a seed word on both sides. So "read the true word" and "read the false word" both fail as one rule applied the same way to every piece. This agrees with what D39, D41 and D60 found by other routes. The practical lesson is: **the lie marks the piece; the word still has to come from somewhere uniform, such as the essence or a pointer.** Ideas 1, 2 and 6 below use that.

### A2. Which lying labels can be decided from data, and which need the eye

- **Decided from data alone (one source, no eye needed):** a title that uses one of the collection's standard label words while the piece's own column says something else. Example: #474 is titled "Invasive ghostGAN" but its species column says bitGAN. This is the basis of Idea 1.
- **Decided from data alone (two sources):** the sheet-versus-blockchain word swaps (#17, #39, #44, #66, #173, #281, #509). These are known and have been run many times.
- **Decided from data alone (label plus number):** a skull whose Type says Glitch but whose number can never glitch under Pindar's own prime rule. This is the basis of Idea 3.
- **Needs the eye:** "Smoke but it is a ghost" (#173), "Rainbow with no greens or blues" (#66, #423), "Silver that looks Flaming" (#81), "Shy that never hides" (#357), "cronenGAN that is not huge or malformed" (#502), "seven eyes, never more than five at once" (#290), "twin that shows three" (#73). This is the basis of Idea 6.

### A3. Where the instruction words sit

Three instruction words sit on pieces that truly glitch (the file shows a returning frame): **adjust** #29, **flip** #142, **jump** #152. All three are order-changing words.
Two instruction words sit on pieces named Glitch that do not glitch: **adapt** #25 and **alter** #60. Both are "change" words.
Two of Pindar's five deliberately wrong hidden-type letters sit on order-changing instruction words: **advance** #32 (its trait reads "32" instead of a letter) and **mirror** #187 (letter C where the rule gives E).
Inside the solution block, **solve** #453 and **sort** #494 are instruction words; **erase** #495 is the piece right after "sort", and **divide** #449 is the piece right before "solution" #450.

### A4. A structural fact I have not seen recorded: the robo drop in a mirror

Pindar gave the robo drop a hidden-type letter F for binary palindromes (#129, #141, #147, #159, #162, #174, #180, #192). To find palindromes he had to read each six-digit binary number backwards. If you do the same to every robo, the pieces fall into mirror pairs. Several named oddities turn out to be mirror partners:

- **#142 "flip"** (001101) mirrors to **#173** (101100), the collector's "on purpose" Smoke-or-Ghost piece.
- **#152 "jump"** (010111), a backwards-Fibonacci glitcher, mirrors to **#187 "mirror"** (111010), a skull/robo hybrid and a deliberate hidden-type misfile.
- **#158 "that"** (011101) mirrors to **#175 "skull"** (101110), a hybrid that sits in two of the collector's pairs.
- **#160 "slender"** (011111) mirrors to **#191 "err0r"** (111110), a hybrid and one of only three essences that is not a plain word.
- **#148 "hint"** (010011) mirrors to **#179 "under"** (110010), one of the four embedded-alphanumeric pieces.
- **#130 "quantum"** (000001) mirrors to **#161 "news"** (100000), the fourth hybrid and another deliberate misfile.
- **#144 "flame"** mirrors to **#189 "blue"**, one of the eleven out-of-sequence seed words.

All four skull/robo hybrids in the robo drop (#161, #175, #187, #191) are mirror partners of #130 or of a glitching piece. Caution, stated honestly: numbers above 32 that end in 0 mirror to odd numbers below 32, and many of those are prime, so some of this would happen by chance (I estimate about one chance in four for the three hybrids). The words "flip" and "mirror" landing exactly on this structure is the part that is hard to call chance.

---

## Part B. Six ideas, ranked

### Idea 1 (best). Eleven titles that lie about their own columns, ending in "final"

- **(a) SELECTION — all DATA, from Pindar's sheet alone.** "A glitch is a piece whose title uses one of my standard label words — a species, a Type or a SubType that other pieces really carry — while that piece's own column says something different." The rule picks exactly eleven: **#2, #17, #154, #293, #474, #478, #498, #502, #503, #509, #512.**
- **(b) WORD.** "The word is the piece's essence." That gives: ability, acoustic, smoke, brand, truth, recycle, safe, source, broken, network, final.
- **(c) ORDER.** "Piece-number order," which for these eleven is also mint order. "final" (#512) falls last by piece number, by mint slot and by the date the file was made.
- **(d) SHAPE.** Eleven words plus a random twelfth. A second, equally plain shape: leave out the Named piece #2 "The Artist" (its title is a proper name, not a label), treat #512 as the bookend it already is, and read **universe + nine (#17, #154, #293, #474, #478, #498, #502, #503, #509) + final + random twelfth**.
- **What in the data teaches it.** The contradictions, each inside one row of the sheet: #2 "The Artist" has Type Glitch (Artist is the Type of seven other pieces); #17 "Nebulous **Ghost** skullGAN" has Type Glitch; #154 "Smokebot **CyberGAN**" has species roboGAN; #293 "Vivid **Glitch** bitGAN" has Type Cloud; #474 "Invasive **ghostGAN**" has species bitGAN; #478 "**Rainbow** bitGAN" has SubType Slender; #498 "Monster **cryptoGAN**" has species monsterGAN; #502 "Inverted **cronenGAN**" has species skullGAN; #503 "Inverted **glitchGAN**" has species skullGAN; #509 "Dark **Rainbow**" has SubType Inverted; #512 "Quantum **glitchGAN**" has species xxxxxGAN. Pindar himself placed "truth" (#474) on one of these and "FALSE" (#473) on the piece beside it. #17 and #509 are already on the collector's "on purpose" list for other reasons. The same rule run on the blockchain records alone returns fifteen pieces: it loses #17 (the blockchain title says Glitch, which matches) and gains #48, #58, #66, #88 and #281. Four of those five are pieces Pindar or the collector already pointed at (#88 "skulGAN" is the spelling slip the collector called an accident), so the rule finds known glitches without being told about them.
- **How stable the rule is.** I counted Twin and Twins as one word and corrected "Gltich" and "Raibow", which matches the collector's taste that spelling slips do not count. With that, the rule returns the same eleven whether "other pieces carry it" means at least two pieces or at least three. If a single other piece is enough, #466 "Invading spookyGAN" joins as a twelfth (only #328 has SubType Invading). Without the spelling corrections, #73, #298 and #379 join. Invented species words (cyclopsGAN, taoGAN, McTestGAN and so on) never count, because no piece carries them in its species column.
- **Already tested?** No. The eleven-word key and both bookend keys are absent from `_registry.json`. The closest earlier run was a union called "Error-named + Name vs GANg mismatch", which shares six words. The tag "Name says one species, GANg says another" has 19 pieces and mixes in the invented tribute species; it was never run as this eleven.
- **What it predicts without a wallet test.** The collector looks at the seven he has not ruled on (#2, #154, #293, #474, #478, #498, #503) and asks one question each: does the picture side with the title word or with the column word? His notes already lean this way for two of them: #502 "is not huge or malformed, which is standard for Cronenberg pieces" (the title lies), and #474 "disappears from frames as Ghost types do" (the title is true and the column lies). If most of the seven show one visibly untrue word, the class is real.
- **Test size.** About 12 phrases: two shapes, each in piece order, date-made order, wordlist order, and those reversed. About 30 if the #466 boundary case is added.
- **Why Pindar would do it this way.** It is exactly "glitch bitGANs with intentional errors": the error is visible in his own sheet, needs no second source, and can be read objectively, like the words in his two earlier puzzles. The order is mint order, as in both earlier puzzles. The phrase ends on "final", the piece he made last.
- **Weakest point.** The rule has two small settings (how many other pieces must carry the word; whether spelling slips count). D78 measured that about 14 of 1,100 simple rules pick exactly eleven, so an exact eleven is weak evidence by itself. #2 is a marginal member.

### Idea 2. Mirror the binary: nine robo glitches point at nine other pieces

- **(a) SELECTION — all DATA.** "The glitches are the robo pieces on prime binary numbers. Read each binary number backwards. Six of them land on each other and cancel (#166 with #170, #172 with #182, #176 with #190). Nine land on a piece that does not glitch: #136, #140, #142, #146, #148, #152, #158, #160, #188."
- **(b) WORD.** "The word is the essence of the piece you land on." In the order of the glitching pieces: #185 digital, #181 vault, #173 worry, #163 time, #179 under, #187 mirror, #175 skull, #191 error (printed "err0r"; Pindar's own index column decodes it as "error"), #184 warm.
- **(c) ORDER.** "The order of the glitching pieces," which is also their mint order. Second choice: the order of the landing pieces.
- **(d) SHAPE.** Universe + nine + final + random twelfth.
- **What in the data teaches it.** Part A4 above. In short: "flip" sits on glitching #142 and flips to the on-purpose piece #173; "mirror" sits on #187, a deliberate hidden-type misfile, and is the mirror image of glitching #152 "jump"; "hint" (#148) mirrors to an embedded-alphanumeric piece; the hidden-type letter F proves Pindar read these binary numbers backwards himself. #142 is also the only robo that glitches on a Fibonacci prime (13), which singles out the word "flip".
- **A second version, stated separately so nothing is mixed.** Selection = the ten robo members of the collector's backwards-Fibonacci eleven (#148, #152, #158, #160, #166, #170, #172, #176, #182, #190). Word = essence of the mirror piece: under, mirror, skull, error, quiz, pink, virus, primary, robot, three. Shape = ten + final, or universe + ten. (The eleventh member, #287, has no six-digit binary number, so it is left out.)
- **Already tested?** The mirror words: no (none of the five keys I built is in the registry, and six-digit binary reversal appears nowhere in DECISIONS.md; only an eleven-digit reversal of the essence index was tried). The nine glitching pieces' OWN words with both bookends: yes, already run with no match (`bookend_t2`), so that plain version is closed.
- **What it predicts without a wallet test.** Show the collector the nine mirror pairs side by side (#142 with #173, #152 with #187, #158 with #175, #160 with #191, #148 with #179, #136 with #185, #140 with #181, #146 with #163, #188 with #184). If the partner is a mirrored or echoed version of the glitching figure in even a few pairs, the mirror is real. A cheap number check: whether the four-in-four hybrid result survives when the same test is run on the spooky drop's octal numbers.
- **Test size.** About 16 phrases for both versions.
- **Why Pindar would do it this way.** It is a program, and the instructions are printed on the pieces: "flip" and "mirror" say what to do, the binary titles say what to do it to, and the palindrome nagas (Evil Olive, Taco Cat) show the move in plain sight.
- **Weakest point.** Nine rather than eleven, so it depends on the bookend shape. Part of the hybrid pattern is expected by chance. "err0r" must be read as "error".

### Idea 3. "Some glitch bitGANs with intentional errors": Glitch on an even number can never be true

- **(a) SELECTION — all DATA.** "In the skull drop a piece glitches only on a prime number. An even number is never prime. So every skull whose Type says Glitch and whose number is even carries a label that cannot be true." There are exactly ten: **#46, #50, #56, #58, #60, #72, #84, #86, #90, #102.** All ten carry Pindar's hidden-type letter "e" (even), and none has a returning frame in its file.
- **(b) WORD.** "The essence": aisle, alert, alone, already, alter, angry, apology, apple, arctic, arrow.
- **(c) ORDER.** "Piece-number order," which for skulls is also wordlist order and mint order.
- **(d) SHAPE.** Ten + final + random twelfth, or universe + ten + random twelfth. A visual twin of this idea, kept separate: the collector's own nine even liars (the same list without #50, which he says merges instead of glitching) as universe + nine + final.
- **What in the data teaches it.** Pindar's 1 January 2022 reply about #58 reads, word for word, "There are some glitch bitGANs with intentional errors — you will find others." #58 is one of the ten: Type Glitch, even number, no frame glitch. On the blockchain Pindar removed the word Glitch from #58's Type, which is the lie being corrected. "alter" (#60) is an instruction word sitting inside this set.
- **Already tested?** Partly. Every eleven of the collector's seventeen liars was run with essences in the plain shape (57 million phrases, closed). The bookend shapes were not: the three keys above are absent from the registry, and #50 was never in the seventeen.
- **What it predicts without a wallet test.** The collector confirms by eye that #50 shows no returning frame. The file measurement already says so.
- **Test size.** About 8 phrases.
- **Why Pindar would do it this way.** It reuses his own number classes from journal 002 and his own hidden-type letter, and it turns the one piece he ever pointed at into a member of a clean class.
- **Weakest point.** It only adds a new shape to a selection whose plain essence reading is already closed. Choosing "even" over "odd" or "square" liars is a choice, even if a principled one.

### Idea 4 (a LESSON about order). Spell the phrase the way Pindar misspells his words

- **What it is.** Not a selection. It is an ordering program that applies to any eleven.
- **(a) SELECTION.** Whichever eleven is under test. I would start with Idea 1's eleven and the three lists the collector marked "maybe" in D90. All are data lists, so nothing is mixed.
- **(b) WORD.** The essence, unchanged.
- **(c) ORDER.** "Put the eleven in mint order, then make the same slips I make in my words: swap the 3rd and 4th, as in Gltich; swap the 7th and 8th, as in Primoridal." Three versions: the first swap only, the second only, both.
- **(d) SHAPE.** Eleven + random twelfth. An extension worth one extra line in the test: "Raibow" drops its 4th letter, so drop the 4th word and end with "final" (ten + final).
- **What in the data teaches it.** "Gltich" is Glitch with letters 3 and 4 swapped, and it occurs on exactly two pieces (#9, #73). "Primoridal" is Primordial with letters 7 and 8 swapped, also on exactly two pieces (#16, #100). Each slip is made twice, the same way, which is what a worked example looks like. The essence "swap" (#402) names the move, and #62 shows the same move one level up (Type and SubType swapped).
- **Already tested?** No. Earlier runs used plain orders, Fibonacci walks, and the solution block as a reordering key. Swapping places 3–4 and 7–8 has not been tried. The word SETS are the same as sets already in the registry, so the registry will say "tested"; only the order is new.
- **What it predicts without a wallet test.** Nothing strong. One check: whether any other double-made slip exists that would add a third swap. I found none beyond "Raibow" (a dropped letter, not a swap).
- **Test size.** About 30 phrases on the four lists named above; about 700 if applied to all seventeen earlier selections and all seven word rules.
- **Why Pindar would do it this way.** "Out of sequence" is his theme at every level (frames, letters, trait fields, essences). This makes the phrase itself out of sequence in the same two places his words are.
- **Weakest point.** The collector ruled swapped letters "look random" and marked both pairs "not sure". It also does not say which eleven to apply it to.

### Idea 5. The solution block as a program: the instruction words are used up, nine words remain

- **(a) SELECTION — all DATA.** "The twelve so-words are the text. 'solution' is the heading, 'solve' is the command and 'sort' is the ordering instruction; once they are obeyed they are used up." Nine remain: #458 soup, #463 sorry, #471 someone, #479 song, #484 soon, #487 soul, #499 sound, #502 source, #508 south.
- **(b) WORD.** "The essence."
- **(c) ORDER.** "'sort' means wordlist order": someone, song, soon, sorry, soul, sound, soup, source, south. Piece order as the second choice.
- **(d) SHAPE.** Universe + nine + final + random twelfth.
- **What in the data teaches it.** "solve" and "sort" are both on the instruction-word list and both inside the block. The only three so-words out of wordlist order in piece order are soup, sorry and sort, so the block visibly needs sorting, and "sort" is one of the misplaced words. #495 "erase" is the piece directly after #494 "sort", in the same working group with the same 89 frames, which reads as "sort, then erase". #449 "divide" sits directly before #450 "solution".
- **Already tested?** No. The registry has the twelve, every eleven of the twelve, and ten-plus-final versions, but not this key: final someone song soon sorry soul sound soup source south universe.
- **What it predicts without a wallet test.** Nothing new to look at; it is a cheap closing test on the block.
- **Test size.** 4 phrases (wordlist order, piece order, and both reversed). Up to 12 if the collector's swap rule is added for #484 (soon becomes exact) and #450.
- **Why Pindar would do it this way.** It is the smallest possible program, and it uses both bookends he gave us.
- **Weakest point.** The block has been run many ways with no match, and deciding that exactly these three words are "instructions" is a judgment.

### Idea 6. The visual version: labels his eye says are untrue (needs a review page first)

- **(a) SELECTION — all VISUAL.** "A glitch is a piece where one label word is plainly untrue of what you see, and the untruth belongs to that one piece." The seventeen Glitch-named skulls are left out because their lie follows from the number rule, not from the single piece. From the collector's notes so far, eleven qualify: #17 (glitches more than it ghosts; not very nebulous), #58 (no frame glitch), #66 (no greens or blues), #73 (says twin, shows three), #81 (looks Flaming, not Silver), #173 (a ghost, not a smoke), #281 (does not glitch), #290 (seven eyes, never more than five at once), #357 (a Shy that never hides), #423 (a Rainbow that is blue and green), #502 (a cronenGAN that is not huge or malformed).
- **(b) WORD.** "The essence": acoustic, already, amused, animal, anxiety, worry, book, boy, seed, cruel, source.
- **(c) ORDER.** "Piece-number order" (also mint order for these).
- **(d) SHAPE.** Eleven + random twelfth.
- **What in the data teaches it.** Every member comes from a collector note in the ledger (D4, D85, D90). I did not add any by my own eye.
- **Already tested?** No; this eleven-word key is not in the registry. The older "mislabel eleven" (D4) was a machine-picked list that the collector mostly rejected; only #357 and #423 survive from it.
- **What it predicts without a wallet test.** This list is eleven only because his review is unfinished, so the count should not be trusted yet. The useful next step is a "Lying Labels" review page: these eleven plus the open candidates (#150, #204, #474, #484, #508, #511 and Idea 1's unreviewed pieces), one plain question each: "Is one label word untrue of this picture, yes or no?"
- **Test size.** About 8 phrases once he has ruled.
- **Why Pindar would do it this way.** It is the collector's own new reading, taken literally, and it needs only watching.
- **Weakest point.** The membership is soft and came from notes written for other purposes. #357 was retired as an anchor in D66. #290's lie is a number, not a word.

---

## Part C. Summary of registry checks

| Word set | In the registry? |
|---|---|
| Idea 1, eleven ending in final | No |
| Idea 1, universe + nine + final | No |
| Idea 2, nine mirror words + both bookends | No |
| Idea 2, ten mirrored backwards-Fibonacci words + final, or + universe | No |
| Nine unpaired robo glitches, their OWN words + both bookends | **Yes, already run, no match** |
| Idea 3, ten even liars + final, or + universe | No |
| Idea 3, nine even liars + both bookends | No |
| Eleven of the seventeen liars, plain shape | **Yes, all 12,376 run, closed (card 57)** |
| Idea 5, nine so-words + both bookends | No |
| Eleven so-words; ten so-words + final | **Yes, already run** |
| Idea 6, eleven visual liars | No |

The registry is a partial record (see `REGISTRY_POLICY.md`), so "No" means "no record found", not proof.