# GAN culture & bitGAN lore — the expanded catalog with test designs

GAN-Culture & Lore Cartographer, 2026-08-26. This file supersedes the thin pass in
`gan_lore_theories.md` (whose CORRECTIONS remain binding: sprite-naming demoted,
painted-words dead, `noohp` is a solver not an architecture) and extends — without
repeating — the 40-item `brainstorm_fodder.md`. Canon checked this session:
`gan_lore_theories.md`, `journal_reread_2026-08-26.md`, `brainstorm_fodder.md` (incl. its
DEAD ENDS appendix, which is honored throughout), `confirmed_glitch_registry.md`,
`WORKER-SUMMARY-2026-08-26.md`, `curated_elevens.md`, `forge_mapping.md`. All 109 pages in
`site/txt/` were read in full this session; every Pindar quote below was taken directly
from those files. Data claims were verified against `bitgan_properties.csv`,
`xmp_timeline.json`, `burnt_final.json`, and `chain_mismatches.json` this session; every
concrete word-set produced below was checked against `theories/_registry.json` (81,662
tested sets) and its status is stated. **No sweeps were run and none are proposed for
autonomous execution — every test is a human-executable design** (per the standing
ask-before-test-runs rule).

Standing constraints every idea must survive: solvable and backtested; spans all 512;
"they have been **looking for glitches all wrong**"; "even knowing how to solve it, it was
hard" (fiddly assembly, not a spreadsheet filter).

Format per item: **{name}** — definition/source → why it might be the key → **TEST**
(concrete, human-executable steps) → word-set + registry status where one exists.

---

## SECTION A — GAN / AI-art history and technical culture

### A1. The adversarial game itself (generator vs discriminator)
**Definition.** A GAN is two networks in a contest: the generator fabricates candidates,
the discriminator classifies each one real-or-fake, and training is the loop where each
gets better at fooling / catching the other (Goodfellow 2014).
**Tie to Pindar.** He never uses the jargon (grep = 0 hits, `gan_lore_theories.md` §D),
but he stages the game constantly: the bitGANs fabricate ("minted itself", "created
themselves in the likeness of four more cryptoArtists", `news__cryptoartgans.txt`) and the
bitGANg's role in every prior puzzle was to *classify* — find the 14 duplicates, spot
which podGANs are out of sequential order, separate "12 that matter" from "2 meant to be a
distraction" (`news__podgallery-solution.txt`). And his provenance taxonomy is a
discriminator's label set: "Some intentional, some natural, and many… I didn't even
intend" (`news__pattern-search.txt`).
**Why it might be the key.** The winning move may be a *classification* act, not a lookup:
the 11 "glitches" could be defined as the pieces that FOOLED the discriminator — anomalies
with no visible tell, only a byte-level one. The perfect fake is invisible; "looking for
glitches all wrong" = looking at the visible glitchers when the real ones show nothing.
**TEST.** Build the discriminator sheet: one row per piece, columns = {visibly glitches
(the collector's eye / visual-worker census), byte-anomalous (re-save, timing, structural),
provenance class (born / planted / natural)}. Then count the "perfect fakes": pieces that
are byte-touched but visually clean. Known members: #47 (silent re-save, the retired
rule-breaker), #203, #407, #454, #457, #460, #461 (re-saved-with-no-detectable-change,
`confirmed_glitch_registry.md`) — that is 7; complete the census from `xmp_timeline.json`
re-saves minus the visible-glitch list and see whether the honest count is 11. If it is,
that set IS a candidate eleven no 2022 solver could have seen.
**Word-set.** None until the census closes; any resulting 11 must be registry-checked
before proposal.

### A2. Latent space, and the walk's PATH as the payload
**Definition.** Every GAN image is a point in latent space; an animation is a walk between
points, and the artist chooses the path. The bitGAN GIFs are visibly latent walks — "They
were changing and a repeating amorphous dance… like looking at a picture that was trying
to come into focus, but couldn't" (`news__the-reveal.txt`).
**Tie to Pindar.** He narrates path properties as meaningful: "They move in a pattern, but
I can't figure out what the pattern is yet" (`journal__page001.txt`); "their movement
appears random but I can't help but wonder if there is a pattern behind it, a dance. After
all they were born from the patterns of a neural net, so it must be deterministic"
(`news__the-dance.txt`).
**Why it might be the key.** If each frame is a latent point, then the *shape* of the walk
(forward, reversed, looped back, held) is an authored channel independent of frame
content — and every confirmed glitch is a path deformity: splices re-visit a point,
holds park on a point, the Review cluster #457/458/459/462/463/497 walks BACKWARD
(`WORKER-SUMMARY-2026-08-26.md`).
**TEST.** A walk-shape taxonomy over all 512, assembled from held data (frame order +
delays, no new sweeps): classify each animation as {plain loop, reversed, palindromic,
splice-revisit, held-point, frozen}. Then ask the two questions never asked: (1) is the
count of non-plain walkers meaningful (is it 11? is it 11 after excluding the born-glitched
robos?); (2) within each deformed walk, WHICH latent point is distinguished (the revisited
frame index, the held index, the reversal pivot) — and hand that index list to the
number-reading tests in B11. Note the dead fence: the speed/timing eleven and
one-frame-reveal eleven straight reads are swept dead; this test differs by using the walk
CLASS as selector and the distinguished INDEX as payload carrier, a combination not in the
registry.
**Word-set.** None yet (indices, not words, are the output).

### A3. Seeds, seed numbers, and the checksum truth about "a random word"
**Definition.** In GAN practice a seed number deterministically fixes the output. In
BIP-39, the 12th word of a seed phrase is not free: it carries a checksum, and for any
fixed first-11 exactly 128 of the 2048 words are legal.
**Tie to Pindar.** The pun is native — podGANs carry a literal `Seed` trait whose value is
a WORD (`journal__page010.txt` §5), closed as a register but alive as a design philosophy:
in his universe *a seed is a word attached to a piece*. And he engineered randomness for
real elsewhere: "6 sets of Random Quantum Data from ibmq_armonk: Each bit was taken by
reading value of a quantum dot in perfect superposition" (`news__quantum-bitgan-collabs.txt`);
"we only want to use each random set once."
**Why it might be the key.** "Ending with a random word" is most economically read as an
honest design statement: the 12th word was chosen at random (possibly literally by his
quantum RNG — the artist who wrote "My art has gone full Quantum… There will be elements
from a quantum computer in every piece of art I have going out", `news__2022-bitgan-roadmap.txt`)
and therefore CANNOT be derived. The clue then tells the solver the solve is 11-words-plus-
a-128-sweep, exactly the shape the sweeps have assumed. The wink pieces (#468 Essence
`random`, #512 Essence `final` with Dimensions `11` — re-verified in CSV this session)
label the sentence, not the mechanism.
**TEST.** No new derivation — a discipline: (1) every future 11-candidate MUST be swept
over all 128 legal checksum words, never just the literal word `random` (already dead) or
a dictionary guess; (2) log, for each past NOHIT, whether the 128-sweep was actually run
(audit `theories/_registry.json` sweep provenance — some early sets were tested with
literal endings only per the dead list). A candidate killed without its 128-sweep is not
actually dead.
**Word-set.** N/A (a protocol, not a set).

### A4. Truncation / temperature — the pull toward the average
**Definition.** StyleGAN's truncation psi drags latents toward the mean; at psi=0 every
seed renders the same average image. Low temperature = less variation = stillness.
**Tie to Pindar.** His glitch origin story is a stillness report: "Lots of my robot's AI
is running, but **it's not moving**" (`news__a-glitch.txt`). The stanza "losing their
ability to move" (`news__text-messages.txt`) has its byte counterpart in the 0ms/frozen
family.
**Why it might be the key.** "Truncated" pieces — animations that barely move, or park on
their own average — are glitches by his own first definition, and stillness is a property
a 2022 solver could SEE without tools, satisfying "hard even knowing how" (you must watch
512 animations attentively).
**TEST.** A stillness ranking from held frames (no new fetches): compute per-piece mean
inter-frame pixel distance; hand the collector the bottom 30 as a review pool ("which of these
truly don't move?"). Cross the resulting still-set against drop conventions (Frozen
SubTypes are supposed to be still — family behavior, not glitch, per the Solstice
color-event lesson) and count the convention-VIOLATING still pieces. If a piece is still in
a drop whose family moves, it is a truncation glitch. Check the final count for 11.
**Word-set.** None until the census closes.

### A5. Epochs and training progression — the drops as a training run
**Definition.** Models improve across epochs; artists exhibit checkpoints as a timeline.
**Tie to Pindar.** He narrates the collection exactly as a training progression: "in each,
the bitGANs discovered new tricks and methods to perfect their form"
(`journal__page016.txt`); "there also appears to be lots of new types, and **new ways of
glitching**" (`news__the-winter-solstice.txt`); "the bitGANs learned to master color"
(`news__the-final-bitgan-drop.txt`). The per-drop glitch-medium map (frames → leet → names
→ inversion → listing → timing) is byte-confirmed curriculum.
**Why it might be the key.** If drops are epochs, positions WITHIN a drop are the stable
coordinate — and the two promise pieces already repeat one position: #32 holds `32` at
skull position 32, #161 holds `32` at robo position 32 (BRIEF §3).
**TEST.** The position-32 probe: list the piece at position 32 of every drop that has one
(skulls #32 `advance`; robos #160; solstice #240; ghosts #332; review #476; spooky has
only 16 pieces — note whether position 32 wraps to 0o.. arithmetic or simply absent, which
is itself informative). Inspect those five/six pieces for anomalies (re-save status,
visible glitch, title oddity) before reading any word. Separately, tabulate
position-in-drop for every confirmed/candidate glitch and look for a shared index rule the
union-of-channels sweep (dead) never tried because it worked in absolute numbers.
**Word-set.** Not yet; if a position rule selects 11 pieces, registry-check before use.

### A6. Checkpoints and save events — the ELEVEN re-save sessions ★
**Definition.** Training and editing both leave checkpoint trails; in glitch-art practice
the file's save history is part of the work.
**Tie to Pindar.** The article that NAMES the phenomenon contains exactly two
observations: "it's not moving. **I see files being saved**" (`news__a-glitch.txt`). And
he confessed the method out of character: "I as the artist can change the form and dance
up until the point I burn the contract" (`journal__le-crucible.txt`).
**Why it might be the key — new count, verified this session.** The 77 post-export
Photoshop re-saves distribute over **exactly eleven editing sessions**
(`xmp_timeline.json`, session ids 5, 6, 9, 13, 18, 24, 25, 26, 40, 53, 56; sizes
26+3+1+1+2+1+13+1+16+1+12 = 77). If "a glitch is a save event" (his own definitional
article), then the natural unit of "11 glitches" may be the eleven SESSIONS, not eleven
pieces. This evades the essence-decoy problem only partly (see caveat) but it is the only
place in the byte record where the number eleven appears as an *authored count of save
events*. Three session anchors are individually suggestive: session 6 opens ON #130
(essence `quantum`, the wallet-cracker with the planted "1024" corruption); session 26 is
#300 alone (the unique PSD-lineage anomaly, saved 12/21 — solstice day); session 56 is the
final 12-file session six days before the clue.
**TEST.** (1) Fix the session boundaries with the collector (the clustering threshold is ours,
not Pindar's — vary it and confirm eleven is stable, not an artifact; that stability check
is the first gate). (2) If stable, derive one word per session by each defensible rule:
first-saved piece's register, last-saved piece's register, the session's date read
Solstice-style, the session's size as a wordlist index. (3) Registry-check each; propose
to the collector before any wallet test.
**Word-sets produced and checked this session (essence register, chronological):**
- session-FIRST essences: `absorb quantum that gown belt beef aware brief wonder crazy
  cloud` — **NOT in `_registry.json` (untested)**.
- session-LAST essences: `adjust cat dog gown blame beef brass brief style crazy song` —
  **NOT in `_registry.json` (untested)**. (Note it ends on #479 `song`, a confirmed
  glitcher.)
Caveat honestly stated: essence straight reads are globally disfavored ("their essence is
just a distraction" + the dead-list's blanket), and the final-session ELEVEN (a different
construction: session 56 minus #512) was swept dead. These two sets are nonetheless
genuinely novel constructions and cheap to test when the collector authorizes.

### A7. Overfitting and memorization — the nearest-neighbor audit
**Definition.** GAN-paper hygiene: show each output beside its nearest training image to
prove the model didn't copy. Memorization = an output that IS a training item.
**Tie to Pindar.** "Pulsating in and out of it is what looked like it's training data"
(`news__the-prime-noungan.txt`). The identity family (#14 wearing #75's art; #150/#179
label swap; #390 double mint) is memorization dramatized.
**Why it might be the key.** The binding correction killed sprite-NAMING, not
nearest-neighbor ANALYSIS. The unworked question: which pieces are anomalously close to
*each other* — near-duplicates short of byte-identity — a relation Pindar planted once at
distance zero (#14/#75) and once as a puzzle in another medium ("The first 100 to find the
duplicate podGANs in the image", `news__1000-eth.txt` — he PLANTS duplicates and asks the
community to find them, that is a proven habit).
**TEST.** Using the held embeddings (`visemb.json` / `framevecs.npy` — no new compute):
rank all cross-piece pairs by visual distance; hand the collector the top ~30 non-identical
pairs for eye confirmation ("are these two the same creature?"). Count confirmed
near-duplicate pairs; check the count against 11 and against drop boundaries (a
near-duplicate SPANNING drops would be a reprise mechanism the sibling-level
correspondence hunt, which is dead at piece→piece level, did not cover because it demanded
reprise semantics rather than raw similarity).
**Word-set.** None until pairs are confirmed.

### A8. Mode collapse — the counting rule
**Definition.** The classic GAN failure: many latents collapse to one output; the model's
diversity is a lie.
**Tie to Pindar.** #14/#75 byte-identical; #390 minted twice; "transforming into a
different bitGan" (`news__text-messages.txt`).
**Why it might be the key.** Fodder A4 flagged collapse as an unexplored COUNTING rule.
Sharpened: the clue counts GLITCHES, not pieces. If two pieces collapsed onto one artwork,
is that one glitch or two? The right census may count *distinct anomaly events*: the
#14/#75 pair = 1 event; the #150↔#179 swap = 1 event; the #390 double mint = 1 event; each
timing hold = 1; etc.
**TEST.** Re-count the confirmed + strong-candidate registry as EVENTS rather than pieces
(pairs and trios collapse to one). Do the arithmetic under each defensible event
definition and see which definition lands on exactly 11. If one does, the eleven words
would be read one-per-event — which immediately raises the good question the pair events
force ("which member speaks?") whose answer Pindar demonstrated in the podGALLERY: marker
pairs sat in First and Last squares — first-member-speaks and last-member-speaks are the
two variants to test.
**Word-set.** Deferred to the event census; registry-check required.

### A9. The loss function — read "loss" literally: the census of LOST things
**Definition.** GAN training minimizes a loss; Belamy was signed with the loss formula.
The formula-signature road is a clean negative here (`gan_lore_theories.md` §D).
**Tie to Pindar.** He keeps narrating LOSS as an event: "Logs showed the total number of
bitGANs **decreasing**" (`news__dad-what-is-this.txt`); "shrinking and disappearing"
(`news__text-messages.txt`); "I thought these were **lost** as I couldn't find them…
Verisart showed me I was **looking in the wrong place**… Imagine losing a treasure then
finding them" (`news__22-for-2022.txt`, eleven days before the clue).
**Why it might be the key.** The byte record now contains a real inventory of absences,
and absences are invisible to anyone "looking for glitches" in what is PRESENT — a clean
fit for "looking for glitches all wrong."
**TEST.** Build the complete absence census, one line each with evidence: #14's rainbow
animation (public nowhere), #134 (the only missing roboGAN GIF, `curated_elevens.md`),
#446's missing Arweave file, #390's trait-less slot 519, the ghost-drop 104→404 "Not
Found" joke, the forge's missing 25th piece (pair burned 2021-10-29, no mint —
`forge_mapping.md` §5), the four skipped BIP-39 words community item #33 in
`news__pattern-search.txt` ("four BIP39 words were skipped in the sequential ordering"),
and any Arweave/OpenSea holes in the archive inventories (`missing_arweave.json`,
`missing_gifs.json`, `missing_urls.json` — already held). Count the absences that are
AUTHORED (not archive rot). If the authored-absence count is 11, the word of each absence
is read from the absent thing's own drop channel.
**Word-set.** None yet.

### A10. Conditional GANs and class labels — the mislabel ELEVEN ★
**Definition.** A conditional GAN generates from a class label; a wrong label is the
canonical conditioning bug — the model renders class A while the label says class B.
**Tie to Pindar.** The SubType system is his label vocabulary, and the byte record shows
one proven label swap (#150↔#179, misfit 6.65→1.97 when exchanged) with "NO textual trace
anywhere" (`journal_reread` §10) — authored silently, or the one glitch class he never
narrated because the labels themselves are the medium.
**Why it might be the key — verified count.** The calibrated visual census of
SubType-vs-art contradictions stands at **exactly eleven pieces**: strong = #179, #423,
#231, #511, #509; medium = #204, #150, #490, #478, #92, #491
(`confirmed_glitch_registry.md`). The instrument was validated by recovering #14 as a
top-2 Rainbow outlier. Eleven pieces whose LABEL lies about their ART is the most literal
possible reading of "glitches" that 2022 solvers universally overlooked — everyone stared
at the art and at the metadata, nobody at the CONTRADICTION between them. And #511's
essence is literally `negative`.
**TEST.** (1) the collector eye-confirms each of the 11 ("does the label really contradict the
art?") — the boundary between medium and weak is instrument-set, so the count needs a
human ratification pass, ideally blind (mix in 10 clean controls). (2) If the eleven
survives, order by mint order (the proven default) and read registers. (3) The natural
extraction has a Pindar precedent: puzzle #19 in his own synopsis — "inverting the BIP39
number to find a different word, particularly when it was an Inverted bitGAN" — i.e., for
label-liars the word may be the TRUE class's word, not the stated one (read what the art
IS, not what the label says: ghost, blue, warm…). Map each piece's true-class word to
BIP-39 and registry-check.
**Word-set (essence register, checked this session):** `arena base cruel fire grab knee
negative network recycle under weapon` — **NOT in `_registry.json` (untested)**.
True-class word-set: requires the collector's namings first.

### A11. Architecture-name culture — cycloGAN is CycleGAN
**Definition.** GAN research names its architectures: DCGAN, StyleGAN, BigGAN, CycleGAN,
BiGAN, InfoGAN, WGAN, pix2pix. Artists in the scene pun on them constantly.
**Tie to Pindar.** His byteGAN mutants are architecture puns: **cycloGANs** (CycleGAN),
**bioGANs** (BiGAN), plus g1itchGANs (`journal__le-crucible.txt`, `news__the-xenogans.txt`,
`news__crucible-unbound.txt`). Post-clue, so precedent only — but it proves he names
creatures after real GAN papers.
**Why it might be the key.** If the habit ran earlier, some Act 1 GANg/type names may
encode architecture references that carry information (which pieces are "marked" as
belonging to the machine layer rather than the story layer).
**TEST.** One-page audit: list every distinct GANg value and Type value in the CSV beside
the canonical architecture list; flag matches and near-matches (cryptoGAN/cyberGAN are
generic; qubitGAN, testGAN, glitchGAN are thematic; anything matching a real paper name is
a flag). Then check whether flagged pieces intersect the anomaly registry. Expected yield
is low — bank as a 30-minute census that closes the question either way.
**Word-set.** None expected.

### A12. DeepDream and the dream corpus — dreams are his clue-delivery protocol
**Definition.** DeepDream (2015) made "the network's dreams" the founding metaphor of
neural art; Pindar's whole clue apparatus runs on literal dreams.
**Tie to Pindar (a pattern, assembled this session).** Every dream in the corpus delivered
an actionable puzzle instruction: the Escape dream showed 16 bitGANs in a spiral → found
at SomniumSpace 2078 (`news__the-escape.txt`); the Moirai seized his hand and wrote the
Solstice barcode clue (`journal__page013.txt`); the Pandimensional lucid dream delivered
the Named-are-the-key doctrine and a META-RULE about reading: "never try and read in a
dream… The text would look like words but be a strain to read… his brain was short
circuiting as it tried to generate the text even though it had insufficient data" —
followed by text that "was moving in the exact same way the creature was. Words would come
in and out of existence… He knew better than to concentrate on them… Instead he just tried
to take in the patterns" (`news__pandimensional-skullgan-ch2.txt`).
**Why it might be the key.** That lucid-dream rule is a startlingly direct authorial
statement published two weeks before the first drop: *do not try to READ words in the
moving art; take in the patterns.* It independently corroborates the painted-words
dead-end (the community searched frames for legible words for years; his founding lore
says reading morphing text is the mistake that wakes you) and points extraction toward
pattern-properties (indices, timing, order, class) rather than content.
**TEST.** Compile the four-row dream table (dream → instruction → what it historically
unlocked) as a standing lens; then apply the meta-rule as a filter over the current
candidate list: any candidate requiring "read words/letters in frames" is
dream-rule-inconsistent (matches the visual dead ends); candidates reading pattern
properties are dream-rule-consistent. Use it as a tie-breaker weight, not a veto.
**Word-set.** N/A (a lens).

### A13. CLIP / text-to-image / diffusion — the fence
**Definition.** CLIP-guided and diffusion art displaced GANs from 2021-22 onward.
**Tie to Pindar.** All his diffusion talk is post-clue ("an ambitious attempt to make a
bitGAN Diffusion Model", `news__2022-bitgan-roadmap.txt`; "experimenting with using
Diffusion Models to add a prompt to each frame", `news__quantum-bitgan-collabs.txt`).
**Verdict.** Clean fence: no Act 1 mechanism can depend on tech he adopted after the clue.
**TEST.** None — this item exists so nobody spends an hour here.

### A14. Pindar's own lineage — cloudpainter, AI Imagined Faces, and the brushstroke numbers
**Definition/source.** His robots ("directed by artificially intelligent software that Van
Arman had developed over the past 15 years", `news__the-moirai-chapter-1.txt`); AI
Imagined Faces = 128 robot-painted BTC editions, "the AI used to make these faces is
nearly identical to the algorithms used to create bitGANs"
(`news__superrare-bitgans.txt`).
**Why it might be the key.** Two hooks. (1) The 128 faces were "lost", then found by
"looking in the wrong place", published 11 days pre-clue — his own metaphor bank for the
meta-tweet. (2) The SuperRare titles carry numbers styled as brushstroke counts — "Three
days and a reported **10,101 strokes** later, what emerged was The Cryptographer"
(`journal__page008.txt`) — and 10101 read as binary is **21**, a Fibonacci number, on the
very first forged piece. The other SR numbers (11,400 / 10,366 / 10,781 / 20,051 / 10,051
/ 11,498 / 7,205…) are not all binary-legal, so the pattern is not uniform — but the first
one being a binary Fibonacci reads authored.
**TEST.** Thirty-minute audit of every SuperRare title number in the corpus: which are
binary-legal, what they decode to, whether the decodes form a set (Fibonacci? piece
numbers?). Include The Cryptographer 10,101 (=21), Laurent 7,205, and the 25 forged pieces
in `superrare_bitgans_onchain.json`. If ≥4 decode into one family, escalate; otherwise
close it as numerology noise.
**Word-set.** None expected (numbers, not words) — any decoded piece-number list feeds
existing register machinery and must be registry-checked.

### A15. The Belamy moment (Obvious, Christie's 2018) — the inverted lesson
**Definition.** "Portrait of Edmond de Belamy" sold for $432,500 signed with the GAN loss
formula — the mechanism inscribed as the signature.
**Status.** Direct reading dead (he never signs with formulas/seeds/numbers —
`gan_lore_theories.md` §D). The inverted lesson stands: Pindar's signature channel is
**provenance metadata** (XMP tool strings, save timestamps, contract boundaries), which is
where every planted glitch actually lives.
**TEST.** Folded into A6 and B14 — no separate work. Kept as one line so the Belamy lead
is never reopened as stated.

### A16. The peer canon — Barrat, Klingemann, Ridler, Sarin — and the tribute census
**Definition.** Robbie Barrat (whose open-source art-GAN code Obvious adapted — the
scene's founding authorship fight); Mario Klingemann ("Memories of Passersby I",
perpetual self-generating art, Sotheby's 2019 — and a co-participant with Pindar in the
Verisart "22 for 2022" auction, `news__22-for-2022.txt`, the one documented direct link);
Anna Ridler (the hand-labeled tulip dataset AS the artwork); Helena Sarin (small personal
datasets — "summon them with my own art" is the same practice,
`news__superrare-bitgans.txt`).
**Why it might be the key.** Pindar's collection encodes PEOPLE as pieces, repeatedly:
four cryptoArtGANs imitate Abosch, Hackatao, Killer Acid, Eclectic Method
(`news__sun-dancing.txt`); four more imitate StellaBelle, Colborn, SuperRareRoses,
Brinkman (`news__cryptoartgans.txt` — the clue page itself!); Ghostly noohpGAN and
Glitching mazzoneGAN imitate his two champion SOLVERS (`news__sun-dancing.txt`); #228 is
Kevin aboschGAN. Person-tribute pieces are a curated, cross-drop, author-selected class —
and the prize piece (Magic Internet moneyGAN) is itself one of them.
**TEST.** Census every person-tribute piece across the 512 (title/name/description
tributes; the four SR description tributes are already known). Count them. Then two reads:
(1) if the count is 11 or 12, test the tribute class as the eleven (order by mint, read
registers, registry-check); (2) regardless of count, note which tributes fall on
anomaly-carriers (#228 is the unique 50ms file — the Abosch tribute IS a timing anomaly;
Abosch's famous work is IAMA Coin, and his tribute description mentions it) and ask
the collector whether the tributed ARTIST names/works supply words (e.g., `coin` is BIP-39).
**Word-set.** Deferred to the census; registry-check required.

### A17. On-chain generative-art norms — mint index vs token number
**Definition.** Art Blocks/fxhash culture made the mint EVENT (index, block, seed-at-mint)
first-class, distinct from the token's display number.
**Tie to Pindar.** Both solved seed puzzles ordered by MINT order, stated verbatim
(`news__the-puzzle.txt`, `news__pattern-revealed.txt`); the #390 double mint and the
#24/#30 relist live in mint space, not metadata space; and 114 pieces show
chain-vs-current metadata mismatches (`chain_mismatches.json`, verified count this
session), most trivially formatting (e.g. #420's title changed from "(120/144)" to
"(420/512)" — the dual-numbering joke resolving).
**Why it might be the key.** The mismatch corpus is the diff between what 2022 solvers SAW
and what the chain originally said — the exact terrain of "looking all wrong."
**TEST.** Taxonomize all 114 mismatches into {renumbering housekeeping, formatting,
substantive}. The substantive residue (trait-value changes like #315 Dimensions 4→5) is
the interesting list: count it, intersect with the anomaly registry, and ask whether ANY
substantive mismatch touches a confirmed glitcher. If the substantive count is near 11,
escalate to a candidate; if it is housekeeping, close the channel with a one-line verdict.
**Word-set.** Deferred; registry-check required.

### A18. Glitch-art orthodoxy — databending, datamoshing, "the file is the medium"
**Definition.** The glitch-art tradition corrupts the FILE (databending) or the codec's
motion model (datamoshing); the artwork is the artifact of the format.
**Tie to Pindar.** Every confirmed Act 1 glitch is file-structural (splice, delay table,
negative frames, re-save, double mint) — orthodox glitch art; `gif_structure.md` and the
inscription rescans already censused the GIF containers.
**Why it might be the key / what is left.** One container channel appears un-censused in
the theory docs: per-frame **disposal methods** and local-palette vs global-palette
choices — the GIF fields that control how each frame REPLACES the last (the datamoshing
analog: wrong disposal = frames bleeding through). A wrong-disposal frame would be
invisible in a frame dump but visible in playback — precisely the visible/invisible split
this puzzle keeps exploiting.
**TEST.** Read `gif_structure_sets.json` / `gifstruct/` for whether disposal-method and
palette-scope were captured; if yes, tabulate deviants per drop convention and count; if
no, flag a one-evening extraction for the collector's approval (all files are local; no paid
calls). Deviant count near 11 → escalate.
**Word-set.** Deferred.

### A19. Artbreeder / GANbreeder — children of two parents, and the eye arithmetic
**Definition.** Joel Simon's GANbreeder crossbreeds two latents into a child — the
folk-genetics of GAN culture.
**Tie to Pindar.** The forge is literal crossbreeding, and he published the inheritance
rule with an arithmetic example: "if the two burnt bitGANs had 2 eyes and 3 eyes, the new
SuperRare bitGAN would have 5 eyes… with a preference for the properties of the
non-flaming bitGAN. It is not an unbreakable rule, however, as some mysterious properties
emerge from nowhere. Like The Extra Terrestrial is an Alien type. Where did that come
from?" (`journal__page009.txt`). 2+3=5 is Fibonacci addition performed on traits, in his
own voice.
**Why it might be the key.** The forge table (`burnt_final.json` + the on-chain `Burned
bitGANs` traits, `forge_mapping.md`) is a Pindar-authored piece→piece→piece relation. If
the inheritance rule is exact, then RULE-BREAKING children ("properties from nowhere") are
authored exceptions — his standard marker grammar (the exception IS the signpost).
**TEST.** For all 16 completed forges: tabulate parent traits vs child traits
(eyes/type/subtype/dimensions) from held data; verify the addition rule and the
non-flaming-preference rule; list every violation. The violations are the candidate
markers — check whether violation-children point at Act 1 pieces (via their `Burned
bitGANs` parents) and whether the pointed-at set has meaningful count. Also finish the
loose thread: WHICH parent pair produced no child (the outstanding debt), and whether the
absent child is itself the marker (crosslink A9).
**Word-set.** Deferred to the violation census.

### A20. The 8-bit training corpus — context, with the fence restated
**Definition/source.** "procedurally generated 32x32 pixel pieces of 8bit art with
references to 1) many of the video games I grew up with & 2) cryptoArt Culture including
things like nounsDAO and cryptoPunks" (`news__act-2-the-collaboration.txt`); "There will
be only ever 2048" sizes the collection to the wordlist (`about.txt`).
**Status.** Sprite-NAMING is demoted by binding correction; the 110-word BIP-39 8-bit
inventory stays banked in `gan_lore_theories.md` §C for the contingency that naming ever
revives via a different mechanism. The unredeemed promise — "tutorials explaining the
**secrets behind how the training data was used**" (`about.txt`) — remains the one live
pointer: the eleven may be facts about how pieces were MADE.
**TEST.** Only this: when any generation-layer census (A1, A6, A19) closes, check its
result against the promise's framing before proposing — the winning mechanism should make
that sentence retroactively true. No standalone work.

---

## SECTION B — the collection's own lore and mythology

### B1. The clue sentence and its parse-space, with the clue-page word census (new negative)
**Source.** "The Act 1 Puzzle: The seed phrase is 11 glitches ending with a random word."
(`news__cryptoartgans.txt`, 2022-02-05). Winks verified in CSV: #468 Essence `random`,
#512 Essence `final` + Dimensions `11`.
**New check run this session.** The roboGAN puzzle's mechanism was "the 12 bip39 words…
were the only BIP39 words contained in the clue" (`news__the-puzzle.txt`). Applied to the
Act 1 clue page: the page contains **35** distinct BIP-39 words (`seed, phrase, random,
word, magic, mystery, puzzle, budget, …`) — far too many for the words-in-the-clue
template to bite without a marking rule (the roboGAN words were "In red"; no styling
survives in our text scrape). **Logged as a checked negative at text level**, with one
residue: the page's rendered/image form may carry the marking (red words, styling) our scrape lost.
**TEST.** Pull the original page from web.archive.org and check the clue post for any
visual marking of words (color, bold, links). Ten minutes; closes the last words-in-clue
door. Also record the parse alternatives once, so downstream tests declare which they
assume: (a) 11 glitch-words + 1 random word appended; (b) the phrase "ending with a random
word" describes each glitch (its END frame is the random-looking one); (c) "ending" names
the final LIST position only (the Solstice grammar — last element pinned). Both solved
templates support (a)/(c); (b) is the novel parse the walk tests (A2) can check for free.

### B2. The Moirai — the Fates, the rule of three, and the three glitch functions
**Source.** #1 "The Moirai" (Primordial, essence `abandon`, three skulls); its one speech:
"This has already happened, so don't try to stop it. And even if you do, it does not
matter. **It will happen again.**" (`news__the-moirai-chapter-1.txt`). Greek Moirai:
Clotho SPINS the thread, Lachesis MEASURES it, Atropos CUTS it.
**Why it might be the key.** The three Fates map startlingly well onto the three confirmed
glitch mechanics: splicing/reinsertion (spinning the thread back into itself), timing
delays (measuring — his one decomposed glitch was literally a list of TIME allotments,
"0.5 seconds, 0.3… Fibonacci times cycle", `journal__page004.txt`), and freezes/cuts
(Atropos — the held frame, the 0ms cut, #47's "cut dialect"). If the eleven is assembled
as three triads-plus-two or a 3/4/4 by Fate-function, the Moirai's rule-of-three is the
organizing myth. Redundant confirmation is already his signature (the negative trio states
one message three times).
**TEST.** Classify every confirmed/candidate glitch by Fate-function {spin=reinsert,
measure=delay, cut=freeze/hold}; tabulate counts per function and per drop. Check whether
any principled cut of the table yields 11 with a 3-flavored structure (3+3+5, 3×3+2,
etc.). Separately, complete the thrice-stated census fodder B2 requested: catalog every
message stated exactly three times in lore or bytes (negative trio; three monoliths
rising, `journal__page008.txt`; three symbols skull/squid/robot, `news__the-ides-are-here.txt`;
three Crucible intelligences; "beginning, middle, and end" repeated in both Named
chapters). Anything thrice-stated in the BYTES beyond the negative trio would be a new
marker.
**Word-set.** Structural; none until a cut is chosen.

### B3. Pandimensional #0 — the hinge, and "before the first, after the last" as an
ordering instruction
**Source.** #0: SubType Pandimensional, Dimensions `infinite`, Level `infinite` (the only
non-"1" Level in the CSV — verified this session), essence `universe`, SuperRare 1/1 on a
different contract, absent from the montage. Its SR description: "**before the first,
after the last, and across every possibility** is the Pandimensional skullGAN"
(`news__the-drop.txt`).
**Why it might be the key.** Every Fibonacci-flavored set flips 10↔11 on whether #0
counts. But the description supplies a third option nobody has tested as an ORDERING rule:
#0 belongs at BOTH ends — first AND last. In a 12-slot seed phrase, "before the first,
after the last" could mean #0's word occupies position 12 (after the last) while the
other 11 fill 1–11 — making `universe` the "random word" candidate, or #0 bracketing the
phrase in derivations that use a passphrase slot.
**TEST.** For each surviving 11-candidate that excludes #0: log the variant "append
`universe` as word 12" as a mandatory cheap test alongside the 128-sweep (single
derivation each, no sweep). Also test `universe` as the BIP-39 PASSPHRASE (the "25th
word") on the strongest historical candidates — a channel the registry format suggests was
not systematically used (registry keys are word-sets, not passphrase variants). Flag to
the collector: confirm with the sweep owner whether passphrase-variants were ever run; if not,
that is a cheap new dimension over already-dead sets, starting with the two
never-tested session sets in A6 and the mislabel set in A10.
**Word-set.** Modifier, not a set.

### B4. The Named pieces — "the key to unlocking the full story", and the name register's
honest failure
**Source.** "a select few would be **Named**, and that these named bitGANs would be **the
key to unlocking the full story**" (`news__pandimensional-skullgan-ch2.txt`). The Named:
the 11 Fibonacci pieces (`journal__page003.txt` lists all, "So if the skullGAN's number is
a Fibonacci, it has a unique name") plus the stated exception: "Only one other bitGAN has
a unique name but is not on Fibonacci sequence - #39 the cyclopsGAN. **Not sure where that
fits in.**"
**Register check run this session.** Reading the unique NAMES as words: artist, whale,
spirit, heart, silver, warrior, bean(ie) are BIP-39; Pandimensional, Moirai, Geometer,
blob, shark are NOT. The name register yields at most 7 legal words of 11 — **it cannot
close as a straight set**, which is itself informative: if the Named are the key, the
words come from a relation, not their names, and not their essences (`fibonacci_names`
essence set already in registry, dead).
**Why it still might be the key.** Two independent pre-clue statements converge on this
eleven with wallet vocabulary ("unlock", "key"); it is the only eleven Pindar ever wrote
down; and its RELATIONS remain barely touched: forge history (The Artist #2 was the first
piece ever burned), gallery placement (see B22), owner history (Anonymoux2311 captured
#0 at 21.34 ETH; gobbles/shartoshi69 captured #1 at 7.7 — both amounts Fibonacci-flavored:
21, 34, 7.7 doubled digits… note but do not lean).
**TEST.** Build the Named-relations table: for each of the 12 named pieces (11 + #39),
one row = {forge involvement, burn date, gallery position in every archived arrangement,
first sale price, owner, montage position, re-save status}. Hunt for the ONE column where
the eleven are distinguished and #39 "fits in" (his phrasing dares the solver to place
it). Any column producing words → registry-check.
**Word-set.** `abandon ability able above abstract account aerobic almost arch bean
universe` (essences) — **in registry, dead**. Names — cannot close. Relations — open.

### B5. The HODL eleven, verbatim, and the unredeemed "unlock"
**Source.** "So HODL the Fibonacci bitGANs (0/128), (1/128), (2/128), (3/128), (5/128),
(8/128), (13/128), (21/128), (34/128), (55/128), & (89/128) cause they will **unlock
special things in upcoming drops**" (`news__hidden-types.txt`). Never redeemed in 109
pages.
**Why it might be the key.** Exactly eleven; the verb is "unlock"; the promise's tense
("in upcoming drops") points FORWARD from drop 1 across the whole collection — matching
the clue's all-512 scope. The straight essence read is dead; what is untested is the
*redemption mechanism*: what special thing did holding these eleven unlock? If the answer
is "they select eleven pieces in LATER drops" (each Fibonacci skull picks a partner in a
future drop by a rule — same number, same position-in-drop, forge descent, gallery
adjacency), the word register would be read off the SELECTED pieces, a two-hop structure
the pointer-sweeps (which tested corruption→partner, not HODL→partner) never covered.
**TEST.** Enumerate the four cheapest partner rules and their selected sets from held
data: (i) same absolute number +drop offset (129+n, 209+n, …); (ii) same position-in-drop;
(iii) forge lineage (which later pieces inherit from burned Fibonacci stock); (iv)
montage/gallery adjacency. For each: does the rule produce exactly one partner per
Fibonacci piece across the five later drops? Any rule that closes → build the word list,
registry-check, bring to the collector. Rules (i) and (ii) are one spreadsheet hour.
**Word-set.** Deferred to partner rule; the four straight-register sets are already dead
per registry.

### B6. Hidden Types, the promise pieces, and 32 as the collection's cell size
**Source.** "(32/128) is Type 32 (typo or important?)… (119/128) is only Odd number of
Type E… These are not 'Hidden Typos', but part of the puzzle that will only be solvable
after some more drops. **Hold on to this info for later.**" (`news__hidden-types.txt`).
Full-data rule-breakers: #32, #119, #161, #187; #161 holds the literal `32` at robo
position 32 (BRIEF §3).
**Why it might be the key.** The number 32 saturates the project's PHYSICAL layer: the art
is 32×32 pixels; reneGANs self-organize "into groups of 32 to fit the aspect ratio of all
major digital displays" (`news__64-more-renegades.txt`); podGANs max-pod at 32
(`news__pods.txt`); bitGANg 1001 is "32 unique bitGANs" (`news__22-for-2022.txt`); 32
testGANs; 32 burns in the burn wallet. The repeated `32` in the promise pieces most
plausibly points at the 32-grid — the collection's atomic cell — i.e., at PIXEL/GRID space
as the final indirection layer ("the secrets behind how the training data was used").
**TEST.** Two probes. (1) Position-32: the piece at position 32 of each drop (skulls #32,
robos #160, solstice #240, ghosts #332, review #476; spooky absent) — pull each one's
anomaly status from held data; if ≥3 of 5 are anomaly-carriers, the promise's redemption
is positional and the drop-wise assembly should run through position arithmetic. (2)
Grid-32: for the confirmed glitch frames (splice/held/negative), record WHICH 32-grid cell
the maximal change occupies (a human eyeballs a difference image); if the cells spell
positions (row/col pairs = wordlist indices in base-32: two cells = one index 0–1023,
inside BIP-39 range at ≤2047 with a 2-cell scheme), that is a generation-layer channel no
sweep has touched. Registry-check anything produced.
**Word-set.** Deferred.

### B7. The skull drop's founding pun — a glitch is a save event
**Source.** "Lots of my robot's AI is running, but it's not moving. **I see files being
saved.** … But this one appears to have a glitch?" (`news__a-glitch.txt`).
**Status.** The lore engine behind A6 (the eleven re-save sessions). Kept as a Side-B
entry only to record the drop-story reading: the FIRST drop's story defines the word the
CLUE uses, and the byte record obeys it. All testing lives in A6.

### B8. The robo drop — binary, escape, and corruption clustering on the wallet-crackers
**Source.** roboGANs numbered 000000–111111; the drop lore is a wallet escape; the
mastermind is qubitGAN #130 (essence `quantum`), whose description carries the planted
"1024" where drop-mates say 2048; the puzzle prize was capturing #130 + accomplice
(`news__superrare-bitgans.txt`). New this session: re-save session 6 (2021-09-24, the robo
period) OPENS on #130 — the first file Pindar touched in that session is the
wallet-cracker with the corrupted count.
**Why it might be the key.** Corruptions cluster ON characters whose lore is about
breaking in (panGAN "cracked my seed phrase with ease", `news__hidden-types.txt`; qubitGAN
1024). If the eleven are chosen "in character," the selection rule may be lore-functional:
pieces whose STORY ROLE is crypto-breaking carry the marks.
**TEST.** Build the lore-role census: every piece whose title/name/description/lore-page
assigns it a crypto function (panGAN #0 cracks seeds; qubitGAN #130 cracks wallets; The
Cryptographer; Magic Internet moneyGAN; Count spookyGAN counts; testGANs test;
noohpGAN/mazzoneGAN solve). Count the crypto-role pieces IN the 512; intersect with the
anomaly registry (already: #0 off-contract, #130 corruption+re-save-first, #200 Type
Glitch pun). If the crypto-role census approaches 11, escalate; read registers in mint
order and registry-check.
**Word-set.** Deferred to census.

### B9. Spooky octal — the counting drop and its one glitch
**Source.** 16 pieces numbered 0o0–0o17; "Count spookyGAN (0o7)" is Type=Glitch (the
counting pun); #205 "spookyGan" lowercase typo; #203 the drop's lone silent re-save
(canon). Per the medium map, spooky glitches in NAMES/metadata and should contribute ~1.
**Why it might be the key.** Choosing among {#200, #205, #203} is a live micro-decision in
ANY drop-wise assembly; getting it principled matters beyond one word.
**TEST.** Decide by cross-drop consistency: whichever anomaly TYPE recurs in other drops'
chosen glitches wins (if the eleven are silent re-saves → #203; metadata lies → #200; case
typos → #205). Run the case-anomaly census of B23 first — if "spookyGan"-style case errors
appear in exactly one piece per drop, #205 wins and the case channel IS the mechanism (see
B23, the strongest new lead of this pass). Also note #200's Essence and #203/#205's under
each scenario for the assembly table; registry-check final assemblies only.

### B10. Winter Solstice — the solved drop's leftover register
**Source.** The Date channel was solved (B16 below). Leftovers: the negative-frame trio
#231/#256/#274 (one identical message three times, planted 12/18 with splice #244); the
authored drop-wide color event frames 33–37 (family behavior); "new ways of glitching. Not
sure they have figured everything out about these new forms yet"
(`news__the-winter-solstice.txt`); open item #280 (palette violates its own date).
**Why it might be the key.** The solved drop DOUBLE-carries: its Date register paid out a
12-word phrase AND it still contains an unredeemed inversion register. Pindar's drops
demonstrably carry more than one channel — so Act 1's solution can reuse a drop already
"spent."
**TEST.** The inversion-family sweeps are dead (registry); what remains is the MESSAGE of
the trio: the negative frames' shared positions were read as the trio "stating a thing
thrice," but the thing stated (the frame-position list itself) was only tested as
skull-number marks (union-eleven, dead). Remaining read: the shared positions as a
DATE (the drop's own solved channel — month/day encodings were its language: "4 were
hidden in a date"). One human hour: take the trio's shared frame positions, try them as
month/day/season encodings under the drop's own solved grammar, see if they name a
calendar word. Registry-check any word produced in context of an assembly.

### B11. Ghosts in the Machine — the listing medium and the reveal→identify→NUMBERS grammar
**Source.** "First noohp discovered **a glitch in the portals**… One frame of each testnet
bitGAN reveals mysterious figures, with tentacles… Closer analysis by FomoBaggins revealed
that these were: Green Shimmer testGAN (3), Yellow Shimmer testGAN (9), Violet Shimmer
testGAN (27). TALK2AM looked at the stats and recognized their numbers were cubes… also
noticed the obscure pattern that 81 is next in the sequence and that there were 81 testnet
bitGANs in the drops" (`news__the-final-clue.txt`). Binding correction honored: the
mechanism is reveal → identify WHICH piece → take the NUMBERS.
**Why it might be the key.** This grammar has never been run end-to-end on Act 1's
distinguished frames. The held/splice/negative frames were checked for foreign ART
(dead — every frame is own art) and for sprites (demoted). But the testnet grammar's
middle step is *identify*, and the identification target in-precedent was a PIECE NUMBER
within the same family. The Act 1 analog: the distinguished frame INDEX itself is the
number (f27, f24, splice indices at primes…), and the number-set — not any word-set — is
the message, to be decoded as a second step (cubes → powers of three in the precedent;
Fibonacci or wordlist indices here).
**TEST.** Assemble the complete distinguished-index list from held data: #479→27, #480→24,
montage-frozen {7,19,31,37,73,244}→their frozen frame indices, splices→reinsertion
indices, trio→shared negative positions, #507/#446/#362 candidate frames. Then the
human step: lay the numbers out as ONE sequence in mint order and ask what the sequence IS
(the precedent's answer was "cubes"; candidates here: Fibonacci times-cycle his
journal names, primes, wordlist indices 1–2048). If the numbers resolve to 11 wordlist
indices, the words follow mechanically. Registry-check the result. (Note: numcode-style
sweeps exist in the workdir — before running anything, diff this construction against
`numcode_*` specs to confirm it is genuinely untested; if identical, mark dead and stop.)
**Word-set.** Potentially direct (indices→words); status pending the diff against
`numcode` history.

### B12. Act 1 Review — the reprise drop, radix corruption, and reading backward
**Source.** Drop narrated in explicit chronological order terminating at #512
(`news__the-final-bitgan-drop.txt`); hex 0x20/0x21 where octal is demanded (o→x revaluing
to 32/33 — landing on the promise number); reversal cluster #457/458/459/462/463/497;
clue-day surgery on #445/#446; the sibling-level homage map (`homage_map.json`: Review
pieces' nearest skulls) with piece→piece correspondence dead.
**Why it might be the key.** The Review is the drop the clue shipped WITH; its glitch
medium is timing; and its anomalies all point BACKWARD (reversal, reprise, "review"). If
the eleven are assembled drop-wise, the Review contributes the ORDERING hint: read the
collection backward, or read each drop's contribution at its reprise position.
**TEST.** One structured hour: for each Review anomaly-carrier, tabulate WHAT it points
back at under the three cheap pointers (same position-in-drop in the original drop; the
homage_map nearest-skull; the binary/hex number's decimal value as a piece number). The
o→x pieces #465/#466 → 32/33 already land on promise piece #32 and robo #33-space; check
whether the other carriers' pointers also land on marked pieces. If ≥4 of 6 pointers hit
anomaly-carriers, the Review is a pointer TABLE and the pointed-at set is the candidate
eleven (then read registers, registry-check).
**Word-set.** Deferred.

### B13. podGANs — Seed, ternary, and the missing radix
**Source.** "every podGAN has a unique ternary and Seed. Perhaps the significance of these
will reveal themselves in the future. Like they have in the puzzles" (`journal__page010.txt`);
powers of three run through the testnet arc (3/9/27/81; podGallery numbers 9, 27, 81, 243,
729, `news__podgallery-solution.txt`); "check out the TERNARY property?"
(`news__the-final-clue.txt`).
**Verified negative this session.** Act 1's `Level` column is `1` for all 512 pieces and
`infinite` for #0 only — no base-3 trace in the CSV; the radix family inside the 512 is
binary (robo), octal (spooky), decimal (ghost jokes), hex (review corruption). Ternary
belongs exclusively to the pod/testnet layer, OFF the prize collection.
**Why it still might matter.** The missing radix is a boundary fact: mechanisms that need
base-3 live off-collection and can be deprioritized for Act 1 — a useful fence. The one
live residue: the testGANs INSIDE the 512 (39 pieces with Type=Tester — see B24's
numerology) are the pod-arc's ambassadors; if any Act 1 mechanism borrows the pod grammar,
it will run through them.
**TEST.** Fence + one check: confirm no Act 1 trait value anywhere parses as balanced
ternary (10-minute scan of the CSV strings). Then close the radix question with a
one-paragraph verdict in the assembly notes.

### B14. The forge/burn ritual — status after `forge_mapping.md`, and what remains
**Source.** Mechanics in `journal__page008.txt` ("You need a FLAMING bitGAN (in Property
or Name)… send both… to the burntGANs wallet"); "The 2 burntGANs will remain in the
burntGANs wallet. Something about the ritual binds them"; 32 burns / 16 forges pre-clue;
exactly 11 burnt Flaming skullGANs.
**Status (respecting the prior worker).** The natural eleven here was found and TESTED:
the eleven forged pieces existing on clue day, essences `super rare series ritual level
secret update switch symbol upgrade chief` — **in `_registry.json`, dead**. The
eleven-burnt-Flaming → eleven-words map does not close (only ten forges from eleven
Flamings; one pair unminted).
**What remains untested.** (1) The `crispy` glitch: Burn Out 10,987's essence is the only
non-BIP-39 word in the forged set (`crisp` + stray letter) — a planted glitch OUTSIDE the
512, proof the glitch program crosses contract boundaries; combined with #0 off-contract,
it keeps alive the uncomfortable hypothesis that not all 11 glitches live inside the 512.
(2) The eye arithmetic and inheritance violations (A19). (3) The unminted pair as an
authored absence (A9).
**TEST.** The cross-contract census: list every anomaly on non-Act-1 contracts that
predates the clue (crispy; #0; the Noun GAN Study in the burn wallet — the only non-bitGAN
ever burned; testnet pieces referenced in mainnet lore). Count them. If
(in-512 confirmed glitches) + (cross-contract anomalies) hits exactly 11 under a
principled boundary, the "Act 1 Puzzle" spans the ACT, not the contract — Act 1 the story
includes SuperRare 1/1s and the forge. Bring the boundary question to the collector before any
word reading.

### B15. SOLVED PUZZLE #1 — the roboGAN seed puzzle (Sep 2021), exactly how it fell
**Step by step, from `news__the-puzzle.txt` verbatim.**
1. "The first step of the solution was to realize that the 12 bip39 words used for the
   seed phrase for the hidden wallet, were the only BIP39 words contained in the clue.
   (In red below)" — the words were IN the clue artifact, marked by color.
2. "The second step was to use the **minting order of the bitGANs** to determine the order
   of these 12 words."
3. Solved by noohp "in under 3 hours."
**Template extracted.** Words live in ONE artifact; a marking rule (color) separates
signal from prose; order = mint order; difficulty ≈ hours. Act 1 is calibrated "much
harder than… the one that took a month" — so Act 1's marking rule is subtler than red
text, but the two-step shape (find the marked eleven → mint-order them) is his invariant.
**TEST.** Already partially run (B1's 35-word census kills the literal transfer). The
residue test is B1's archive pull for visual marking. Then retire this template except as
the order-authority it is.

### B16. SOLVED PUZZLE #2 — the Winter Solstice (Jan 2022), exactly how it fell
**Step by step, from `news__pattern-revealed.txt` verbatim.**
1. "Their nature was seasonal, so it was all in the **date properties**." One field.
2. Three hiding depths: "The first 2 seed phrases were in plain site, 'day' and 'night'";
   "Then 4 were hidden in a date. Each of the cryptoArtistGANs had a date that fell on
   either and equinox or solstice and represented 'spring', 'summer', 'autumn', and
   'winter'"; "the next 6 were hidden in plain sight as the 6 months that are also bip39
   words."
3. "And finally, their order was according to **mint order**. So if you lined up all the
   relevant seed words and read down the list you would get the final seed phrase":
   `day spring march april august summer october autumn december winter february night`.
4. The order hint, when one was needed, pinned ONLY the ends: "It begins with sunrise and
   it ends with sunset… everything else is in between" (`news__pattern-search.txt`), and
   the essence-decoy warning was explicit: "Their essence is just a distraction, they hide
   their true BIP39 nature elsewhere."
**Template extracted.** ONE field; a small taxonomy of hiding depths (plain sight /
decode / list-membership); mint order; ends pinned when hinted; essences are the decoy BY
STATED DESIGN. The Act 1 clue pins only the last slot ("ending with a random word") —
exactly this grammar.
**TEST — the field-by-field depth census (the most template-faithful unrun test).** For
each Act 1 field in turn (Title, Name, GANg, SubType, Type, Eyes, Dimensions, Drop):
count words recoverable at each depth — (i) plain-sight BIP-39 values, (ii) decodable
values (dates→seasons analog: numbers→wordlist indices, radix decodes), (iii)
list-membership values (values that are BIP-39 AND anomalous in distribution — see B21).
The winning field, per template, should yield 11 words across 2–3 depths with a clean
story per depth. `name_mining.md`/`word_sources.md` cover pieces of (i); the DEPTH
taxonomy per field has never been tabulated as such. Registry-check any 11 that emerges.

### B17. SOLVED PUZZLE #3 — the podGALLERY (Dec 2021), exactly how it fell
**Step by step, from `news__podgallery-solution.txt` verbatim.**
1. "Stage 1: Find the Duplicates. There are 14 duplicates between the 6 pods - **12 that
   matter, and 2 meant to be a distraction**."
2. "Stage 2: Recognize the Pattern… at least 1 of the duplicate pairs was in the **First**
   and **Last** square of each pod." Multiple routes stated: comparing before/after
   states; "significance in the numbers 9, 27, 81, 243, & 729"; "the first and last
   podGANs were the only ones **out of sequential order**… It didn't matter how you
   discovered the pattern."
3. "Stage 3: … the seed phrases in the attributes are the **wrong words**. They are words
   from the slip-39 list, when the words needed… needed to be taken from the bip-39 list.
   So if you looked up the equivalent value from the bip-39 list, you would got the
   correct seed phrase, in the correct order."
4. "Stage 4: Open the Wallet." Plus: "Sorry for the two red herrings… they were tough.
   This is first time one of my puzzles took longer than 5 minutes to solve."
**Template extracted.** Stated decoy counts; POSITIONAL markers (first/last); multiple
convergent discovery routes; and the final step is an INDIRECTION (right index, wrong
list). The SLIP→BIP hop into Act 1 was swept (zero adherents, dead) — but the ABSTRACT
lesson stands: expect the last step to be "you have the right POSITIONS but are reading
the wrong LIST/register."
**TEST.** Apply "right position, wrong list" to the strongest positional candidates: for
any 11-set already believed positionally correct but word-dead (the union-eleven, the
Fibonacci eleven), enumerate the OTHER lists an index can address (BIP-39 is one of many:
the 512 titles list, the drop-internal orders, the montage order, the upload order in
`act1_uploadorder.json`) and check one hop — index in list A → entity at same index in
list B → that entity's word. This is the podGALLERY move abstracted; the multistep sweeps
covered register-to-register hops on the SAME pieces, not index-transplants BETWEEN
orderings (verify against `multistep_sets.json` before running; anything already there is
dead).
**Word-set.** Deferred; registry-check mandatory.

### B18. The byteGAN puzzle (Mar 2023) — the very next seed puzzle, method precedent
**Step by step, from `news__the-bytegan-puzzle.txt` verbatim.**
1. Unintended route (bitquence): "noticed I was uploading and inscribing GIF images on
   chain, and that **all the images had tokens associated with them except one**. He then
   downloaded that GIF off of the blockchain and examined it. **Within the metadata he
   found Photoshop layers that corresponded with a 12 word seedphrase**."
2. Intended route (noohp, "slightly behind"): "the gif was **12 frames** and **inscribed
   very lightly in the background were 12 seed phrases**."
3. House rule: "it doesn't matter how you find the seedphrase and there is no judge."
**Why it matters.** His NEXT design after Act 1 used: (a) the odd-file-out as pointer, (b)
Photoshop provenance as an unintended-but-honored channel, (c) per-frame payload. All
three shapes now have Act 1 analogs surfaced by the byte work (77 re-saves; one-frame
holds; #446/#134 odd files). The faint-inscription channel specifically was re-scanned in
Act 1 and is dead (the collector rejected the letter residue) — but the ODD-FILE-OUT pointer
grammar has not been run as a chain: odd file → examine THAT file's structure → payload.
**TEST.** List Act 1's odd-files-out by each criterion (the only missing GIF #134; the
only 50ms file #228; the only PSD-derived file #300; the only double mint #390; the only
off-contract piece #0; the only file with a missing Arweave twin #446). For each, one
human-hour of structural examination of THAT file alone (frame count, layer residue, delay
table, palette) asking "does this one file contain a LIST?" — the byteGAN payload was 12
items inside ONE file. A 12-frame or 11-anomaly structure inside a single odd file would
be the byteGAN design one iteration earlier. (The steg sweeps were corpus-wide and
generic; a single-file bespoke read against a specific 11/12-list hypothesis is different
and cheap.)
**Word-set.** None until a file yields structure.

### B19. The montage and every published composite — the planted-duplicate precedent
**Source.** He planted duplicate podGANs in his published 24,000×13,500 poster and made
FINDING THE DUPLICATES the puzzle: "The first 100 to find the duplicate podGANs in the
image above will be sent a signed limited edition postcard" (`news__1000-eth.txt`). Act 1
montage findings to date: display-order permutation at skull positions {22,23,24,29,30};
six tiles frozen on splice frames; #0 absent; every tile own-art (`montage_verdict.md`).
**Why it might be the key.** The 1000-ETH poster proves duplicates-in-composites is a
Pindar puzzle primitive. The Act 1 montage was cleared of foreign art, but was it ever
scanned for REPEATED tiles (same piece appearing twice, or two tiles showing the same
frame of different pieces)? The permutation event shows the montage photographs his
surgeries; a duplicate would photograph the #14/#75-class events.
**TEST.** Tile-vs-tile duplicate scan WITHIN each published composite (all.jpg, bitGANg
1001's 32 tiles, `act1review_8x8.gif`, the finaldrop grid, contact sheets): exact and
near-duplicate tile pairs (held tooling: `dupscan.py`, `montage_pipeline_*` — verify scope
first: if dupscan already covered tile-pairs, mark done and close). Then the same scan on
bitGANg 1001 specifically — 32 pieces CHOSEN by Pindar for the pre-clue Verisart auction
(published 11 days before the clue, `news__22-for-2022.txt`): the selection of which 32,
and whether any anomaly-carriers are among them, has never been read as a signal. List the
32, intersect with the registry.
**Word-set.** Deferred.

### B20. The CryptoVoxels galleries — arrangement-as-password, the strongest spatial precedent
**Source.** The Hidden Journal's password WAS a gallery arrangement: "the password which
was **the order of the NFTs in the gallery [01123581321345589]**"
(`news__hidden-types.txt`) — the eleven materialized pieces hung in Fibonacci order, and
their hanging order was the key. Also: "the best place to hide it is right under the NEXUS
of their activity, the bitGALLERY" (`journal__page001.txt`); spirals with a payload at the
center ("a slightly off color tile at the center of the spiral where I stashed the Hidden
Journal"); "TO FULLY UNDERSTAND YOU MUST GO BEYOND THE END OF THE SPIRAL"
(`news__what-have-we-done.txt`).
**Why it might be the key.** This is the only precedent where ELEVEN pieces + their
DISPLAY ORDER literally formed a key — the exact shape of the Act 1 problem. If Pindar
re-used the move, the Act 1 eleven's ORDER (or identity) was exhibited spatially in a
gallery/parcel arrangement sometime before or at the clue date, and screenshots exist.
**TEST (archive hunt, pure human work).** (1) Enumerate every gallery/parcel he used
(CryptoVoxels bitGALLERY at the quoted coords, META Gallery, SomniumSpace 2078, the
Voxels crucible plot) and every date-stamped visual record (his tweets, Partyvan's
orientation pic he praised, community screenshots, web.archive captures of
cryptovoxels.com pages). (2) For each capture near 2022-02-05: WHICH pieces hung, in WHAT
order. (3) Any arrangement of ~11 Act 1 pieces → read its order against mint order; any
deviation is signal (podGALLERY: "out of sequential order" WAS the marker). This has never
been done for the clue-era galleries; it is slow, external, and exactly the kind of
"fiddly observation" the difficulty quote promises.
**Word-set.** Deferred to a found arrangement.

### B21. "Unevenly distributed" — count-anomaly as a hiding depth
**Source.** The Solstice winner's handle was literally **unevenly_distributed**; community
theory #21 he chose to publish: "the bitGANs are unevenly distributed between the months,
and that the months with fewer bitGANs are BIP39 words" (`news__pattern-search.txt`).
**Why it might be the key.** Distribution-counting is a hiding depth he has flirted with
publishing: the VALUE that is rare (or over-common) within a field is the marked one. It
requires counting all 512 — "hard even knowing how" — and is invisible piece-by-piece.
**TEST.** For each CSV field: frequency table of values; flag values whose count is
anomalous against the field's shape (singletons, or counts hitting meaningful numbers —
11, 32, Fibonacci) AND whose value or whose members' register is BIP-39. Known instances
to calibrate: 8 GANg=glitchGAN pieces with #281 the sole non-glitcher (the 1-of-8
singleton); Eyes=0 exactly 4 pieces (three b0rg cyberGANs + Anonymous spookyGAN — "blind
to everything but data, its eyes themselves are information", `journal__hidden-lore-revealed.txt`,
is the Binary roboGAN's lore, showing eye-count carries meaning). Output: a shortlist of
count-anomalies for the collector. Registry-check any resulting set.

### B22. The Eyes channel — countable lies a human can catch
**Source.** Eyes is a first-class trait (distribution verified this session: 0→4 pieces,
1→6, up to 10→2); the forge ADDS eyes (2+3=5, `journal__page009.txt`); the lore keeps
staring: "So many eyes, all staring into me from angles I can not even see"
(`news__they-materialized.txt`); "look closely into the eyes on the back wall… and was
transported" (`news__the-final-clue.txt` — the eyes were literally the portal to the
hidden message).
**Why it might be the key.** Eyes are the one numeric trait a 2022 solver could VERIFY by
looking — count the eyes in the art, compare to the trait. A piece whose art shows a
different eye count than its Eyes value is a "glitch" that is simultaneously visible,
checkable, and universally overlooked (nobody audits a number that looks decorative).
Anirveda's precedent says eyes are where portals hide.
**TEST.** Human audit, sampling first: the collector counts eyes on (a) all confirmed/candidate
glitchers, (b) the Eyes=0 and Eyes≥7 tails (14 pieces), (c) 20 random controls. If liars
exist beyond noise, extend to the full 512 (the `glitch_liars.json` artifacts in the
workdir — 23 and 17 element lists — appear to be an earlier liar-hunt: reconcile first
with whoever built them; if they ARE eye-liars, the census is half done and its
intersection with the registry is immediate).
**Word-set.** Deferred to the audit; if exactly 11 liars, read registers in mint order and
registry-check.

### B23. The materialization acrostic — a VERIFIED planted letter-code in a published list ★
**New finding, verified this session against `news__materialization.txt`.** The 33 artist
names under the materialization gallery carry displaced capitals that spell the alphabet
IN ORDER: gramm**A**r error, **B**uzz, pinko**C**to, mollieh**D**, d**Ξ**cision (Xi
standing at E's position), **F**omobaggins, g**G**ainor, r**H**ye, d**I**nomons, n**J**c,
stin**K**ydaddy, name**L**essname, thatbloo**M**, pla**N**tman, lam**O**rak,
floo**P**moose, wearys**Q**uirrel, ez**R**a, hump**S**tyles, t_t_**T**_tim,
the**U**rbanita, anir**V**eda, **X**abier, ke**Y**chain, kee**Z** — A through Z with W
absent and eight uncapitalized names trailing (0xinuarashi, ii32x4, eskalexia, 1m3y,
sanguineseal, mc, fams3, bobolo). This is not community content; it is Pindar's own news
page, and it PROVES he plants ordered letter-codes into published lists via case
anomalies.
**Why it might be the key.** Case anomalies are exactly the texture of the known
metadata glitches ("spookyGan" #205; leet robo names; "Gl1itch1ng"). If he encoded an
A-to-Z in an artist list for fun, the same channel in the 512's TITLES could carry the
real payload — and case is invisible to every solver who normalizes strings (which is what
spreadsheets and our own tooling do by default: "looking for glitches all wrong" indeed).
Also note: W missing + 8 trailing names — check whether the absent letter or the
unmarked-8 mean something (the eight uncapitalized include ii32x4, the relay of the
text-messages taxonomy).
**TEST.** (1) Exact-case census of all 512 on-chain ORIGINAL titles/names (from
`chain_raw_by_mintindex.json`, not the CSV, which may be normalized): flag every token
whose capitalization deviates from its family convention (lowercase where siblings
capitalize, mid-word capitals, digit-for-letter swaps). (2) For each deviant: the deviant
LETTER and its POSITION are the payload candidates (the acrostic proves letter+position is
his encoding unit). (3) If ~11 deviants exist, their letters in mint order form a string —
check it as initials of BIP-39 words, as a word itself, and as an ordering key.
(4) Separately, re-read the other published LISTS (the credits list in
`news__almost-there.txt` with its "Van Arman x Connery" style pairs; the hidden-types
solver list; the pattern-search theory list) for the same trick. The word-transform
dead-end (reversals/atbash/Caesar on WORDS) does not cover case-position codes — this
channel is genuinely untested.
**Word-set.** None yet; any derived set → registry-check.

### B24. Numerology audit — 11, 32, 33, 39, 128, 512, 2048, cubes
**Sources, all verified or quoted.** Eleven: the HODL eleven; "So 11 bitGANs have
materialized"; "Another 11 just appeared"; 11 burnt Flaming skulls; 11 re-save sessions
(new, this pass); #512 Dimensions=11; post-clue 1,111 / 11×11 / "11 Crucible Events".
Thirty-nine: #39 cyclopsGAN ("Not sure where that fits in"); "39 manifestations in total"
(`news__materialization.txt`); **Type=Tester count = 39** (verified in CSV this session).
Thirty-three: "33 artists in total"; "a total of 33 new bitGANs". Thirty-two: B6's cell
size. Cubes/powers of three: testGANs 3/9/27, podGallery 9/27/81/243/729, "eight cubes in
a tesseract… 2, 4, 8, 16, 32, 64 - These are powers of two" (`journal__page005.txt` — he
walks the reader through dimensional numerology himself).
**Why it might be the key.** Three independent 39s (a named-exception piece, a
manifestation count, a Type census) is the kind of self-confirming redundancy he builds
(the Moirai habit). If #39 "fits in" anywhere, it fits as the INDEX of the Tester class —
the class whose lore is testing puzzles.
**TEST.** The count-the-counts audit: tabulate every count Pindar states aloud in the 109
pages (11, 14, 12+2, 16, 18, 22, 33, 39, 81, 88, 92, 128…) against the byte record's
actual counts; flag every stated count that MISmatches reality (his "93 new bitGANs" in
`news__sun-dancing.txt` vs 92 elsewhere is one already — decide typo vs signal), and
every count stated twice in different domains (the 39s, the 33s). Mismatches and doubles
are candidate signposts; the specific follow-up for 39: list the 39 Testers, check
anomaly intersections, and ask whether #39-the-piece points at the class (its essence
`age`, its Flaming/1-eye traits) — one hour.
**Word-set.** Deferred.

### B25. The Moirai's dictated clue page and the barcodes
**Source.** "Last night the journal opened and a familiar force took control of my hand.
It was The Moirai (1/128) and it directed me to write down the following. It doesn't make
any sense to me" (`journal__page013.txt` — the written-out content survives only as
images, not text); community decode: "One of the barcodes in the second clue was
deciphered to **6621439** and mapped to a specific bitGAN" (`news__pattern-search.txt`).
**Why it might matter.** The Solstice-era barcode page is spent (its puzzle solved), but
it establishes that Pindar delivers machine-readable codes (barcodes) inside journal
IMAGES — a channel our text scrape cannot see at all. If any Act 1-era page image carries
an unspent code, no text-based pass would ever have found it.
**TEST.** Inventory every image embedded in the 109 pages (from `site/raw/` HTML: img
URLs), date each, and human-inspect the Act 1-window images (Jan–Feb 2022 posts,
`news__cryptoartgans.txt` above all) for codes: barcodes, QR fragments, deliberate pixel
patterns. The clue post's own images are the highest-value targets and have never been
audited as carriers (the letter-projection sweeps ran on the 512 GIFs, not on the WEBSITE
imagery). Cheap, bounded, and closes a real hole.
**Word-set.** None yet.

### B26. "Some intentional, some natural" — the provenance triad as the counting frame
**Source.** "there are a lot of patterns hidden in the bitGANs. Some intentional, some
natural, and many that were discovered by the bitGANg that I didn't even intend"
(`news__pattern-search.txt`). Byte-confirmed: born-glitched robos (export-time) vs planted
skull/solstice glitches (re-save sessions).
**Why it might be the key.** The clue counts eleven GLITCHES. If Pindar counts only the
INTENTIONAL class — and the byte record can now separate intentional (planted in re-save
sessions) from natural (born at export) — then the census that matters is
planted-visible-glitches only, and the robo glitches (born) may NOT count toward the
eleven at all, which would break every census that included them.
**TEST.** Re-run the count arithmetic of A8 under three provenance filters: intentional
only, natural only, both. For each filter × each event-definition, record the count.
Any (filter, definition) pair landing on exactly 11 becomes a named candidate with its
member list; bring the list (not words yet) to the collector for the register decision.
Registry-check follows.

---

## RANKED SHORTLIST — the 12 most promising ideas across both sides

1. **B23 — the case-anomaly channel (materialization A–Z acrostic).** A verified,
   previously unknown demonstration that Pindar plants ordered letter codes in published
   lists via displaced capitals — run the exact-case census on the 512's original on-chain
   titles.
2. **A6 — the eleven re-save sessions.** "A glitch is a save event" + exactly 11 re-save
   sessions covering all 77 touched files; two never-tested word-sets already in hand
   (session-firsts, session-lasts — both absent from the registry).
3. **A10 — the mislabel eleven.** The calibrated SubType-vs-art contradiction census
   stands at exactly 11 pieces, its essence set is untested, and the "read the TRUE class"
   extraction matches Pindar's own published inversion idea.
4. **B20 — gallery-arrangement-as-password.** The Hidden Journal's password literally was
   eleven pieces' display order; nobody has recovered the clue-era gallery arrangements
   from archives.
5. **B16 — the Solstice field-by-field depth census.** His proven grammar (one field,
   three hiding depths, mint order, last slot pinned) has never been systematically
   tabulated across Act 1's fields.
6. **B11 — reveal→identify→NUMBERS run end-to-end.** The testnet grammar's output was a
   number sequence (cubes); assemble Act 1's distinguished frame-index sequence and ask
   what IT is — after diffing against the numcode sweep history.
7. **A1/B26 — the discriminator census (perfect fakes / intentional-only counting).**
   Count byte-touched-but-visually-clean pieces and re-count glitch EVENTS under
   provenance filters; either count landing on 11 names a candidate immediately.
8. **B18 — the odd-file-out bespoke read.** His very next puzzle hid 12 items inside the
   one file without a token; examine Act 1's six odd-files-out individually for internal
   11/12-structure.
9. **B17 — "right position, wrong list" index transplants.** The podGALLERY's final step
   was an indirection; transplant the positionally-strong dead sets' indices across Act
   1's other orderings (upload, montage, title-list) — after diffing against multistep
   history.
10. **B22 — the Eyes lie audit.** The one numeric trait a human can verify against the
    art; the lore says eyes are where portals hide; reconcile with the existing
    `glitch_liars.json` artifacts first.
11. **B6 — the 32 probe.** The promise pieces' repeated `32` read as the collection's
    32-grid cell size: position-32-of-each-drop and grid-cell reads of the glitch frames.
12. **A19/B14 — forge inheritance violations.** Verify the published trait-addition rule
    across all 16 forges; the violations ("properties emerge from nowhere") are authored
    exceptions in his marker grammar, and the unminted 25th piece is an authored absence.

## Counts
- Side A items: **20** (A1–A20).
- Side B items: **26** (B1–B26).
- Items carrying a concrete word-set with registry status stated: **6** — A6 (two sets,
  both untested), A10 (one set, untested), B4 (essences dead / names cannot close), B14
  (forge eleven, tested-dead), B5 (straight registers dead, partner rules open), B16
  (solved-phrase precedent recorded). Three untested word-sets total are ready for
  the collector's go/no-go: session-firsts, session-lasts, mislabel-11 essences — none may be
  swept without his authorization.