# The bitGAN Act 1 puzzle: a first-principles brief for a new run

Written 27 September 2026 for a fresh Claude chat that will lead a swarm of agents. It is self-contained. Read it top to bottom once, then use the data map at the end. Everything it cites is public at https://bitganspuzzlebook.com or in the local repository named in the data map.

## 1. The puzzle in one paragraph

Pindar Van Arman made 512 animated GIF artworks called bitGANs and released them in six drops between August 2021 and August 2022. A wallet, `0x18f87ec9c527aba1db44f715456bf28b0dae478d`, holds the prize. Its 12-word BIP-39 seed phrase is the answer. Pindar's clue: **"The seed phrase is 11 glitches ending with a random word."** Each of the 512 pieces carries an "essence", a single word from the 2,048-word BIP-39 list, so the natural reading is that eleven pieces (the glitches) supply eleven words in some order, and the twelfth word is one of the roughly 128 words that make the checksum valid. The puzzle has been open since February 2022. Nobody has solved it.

## 2. What the collector wants from this run

The collector (Michael, who owns the puzzle-book site) wants a run that starts from first principles rather than from our old candidate lists. His instructions, in his words as far as possible:

- Every drop had its own logic, encoded in how the pieces were ordered, named, or how they look. We now have that logic written down for all six drops, from three sources: the chain data, Pindar's own spreadsheet, and the collector's frame-by-frame visual audit of all 512 pieces (finished 27 September 2026).
- Use the logic of the drops to look for a larger logic of the whole collection. There are 512 pieces. Pindar thinks in bits and bytes. Ask what the logic of organising 512 bits of information is, in the puzzle and in computer programming generally.
- Look for things that break the logic of a drop. Look for things that echo one drop's logic elsewhere in the collection.
- Look at each individual drop as a map to Act 1 Review. Look at Act 1 Review as a map pointing back to the original drops.
- Explore the numeric, visual and other logics of how the collection was put together, and use them to find the eleven glitches that combine with the random word.

### The collector's three tests for any candidate eleven (standing, 21 September 2026)

1. **Cross-drop.** The eleven must come from more than one drop. A set drawn from a single drop is de facto wrong.
2. **Hard even for Pindar.** The selection should be something a determined solver could find but not stumble into.
3. **The last word is random.** The twelfth word is not derived; it is whatever makes the checksum work. Pindar's own words. (One piece, #468 Flaming testGAN, has the essence 'random', and #512 has 'final'; both have been used as literal words in past runs.)

### Standing rules of work

- State the hypothesis before any sweep: which pieces, how each becomes a word, in what order. Then run every variant you can state (all orders, forwards and backwards, and the 'ten + final + random' shape as well as 'eleven + random'). Do not ask permission for variants; run them and report plainly what was tested and why.
- Never mix sides in one set: a set of glitches is all data (chain, sheet, file) or all visual. Never both.
- No machine-vision look-alike matching. Colour measurement is fine as support; the collector's eye decides.
- Write everything in plain sentences. No codes, chips or grids in anything he reads.
- Every finding goes to the ledger, then to the site. Nothing is ever deleted; corrections are appended.
- Money: never spend it, never touch the wallet, never call paid services without his go-ahead.
- Check the registry before running: about 240,000 distinct eleven-word sets have been tested with no hit (each against every legal twelfth word and 33 derivation paths). A miss says that exact combination missed and nothing more, but do not spend time re-running it.

## 3. How to run and record a sweep

The harness lives in `~/Desktop/bitgan_puzzle_work`. Build a task list of (label, eleven words) pairs with the pattern in any `build_hyp*.py` (the helpers `full()`, `shapes()`, `add()` in those files apply seven word rules and seven orders forwards and backwards). Then:

```bash
cd ~/Desktop/bitgan_puzzle_work && python3 build_hypNN.py && for i in 1 2 3 4 5 6 7 8; do nohup python3 qrun2.py sweeps/hypNN $i > sweeps/hypNN_w$i.log 2>&1 & disown; done
```

The builder plants one genuine task's wallet in the queue. A run is valid only if a worker prints `[plant seen]`. A real hit writes `RECOVERED_HITS.txt` and stops. Record the result as a plain-language card with `bitganspuzzlebook/scripts/push_findings.py` (hypotheses and runs), add a D-entry to `theories/DECISIONS.md` (next number D111), and refresh the site with `./update_all.sh` (retry on Cloudflare error 10000) and a Pages deploy. Cards run to 244 so far.

Word rules used so far: the essence; the wordlist word at the piece number, the mint slot, or the place in the drop, each counted from 0 or from 1. Orders used so far: piece number, mint slot, date the file was made, date it first left Pindar's wallet, wordlist place, title A to Z, the counter printed in the title, frame count, calendar date (Solstice), frame number in the filename (Solstice), and Pindar's Review working groups. All of these are on record; the new run should look for orders and word rules that come from the collection's own logic rather than from sorting a column.

## 4. The logic of each drop

This is the core of what is new. Each drop's page on the site carries a "How the drop is laid out" section and a rules list with breaks (`/data/drop_rules.json`). Summary:

### Drop 1, skullGANs (#1 to #128, 20 August 2021)

Numbered 1 to 128 and the number does the work. Pindar sorted every number into one of five classes in a fixed order (his journal page 002): Fibonacci, then the remaining primes, then squares, then remaining odd, then remaining even. Each class has a look: primes glitch (frames jump back in time; the file confirms it); odd non-class numbers are Flaming (26 of 26); squares are Silver or Smoke (9 of 10); Fibonacci numbers have 5 Dimensions (9 of 10). The essences run the wordlist in order (abandon to answer). The chain carries a hidden-type letter recording the number class. Breaks: #16 (a square that is neither Silver nor Smoke); #31, #41, #55, #89 (Silver or Smoke off the squares); #7 and #8 (their Dimensions look swapped; pair P15); #47 and #50 (glitch against the rule); the Eyes trait disagrees with the collector's count on 21 of 128 skulls. Three Named pieces (#1 The Moirai, #2 The Artist, #3 The Geometer) open it.

### Drop 2, robots (#129 to #192, 29 September 2021)

Counted by the 6-bit binary number in the title, 000000 to 111111, and the binary number sets the class: palindromes are the eight nagaGANs (all named with palindromes, all with a colour flash or speed-up by eye); non-Fibonacci primes are Type Glitch (14 of 14); squares from 4 up are Smoke (5 of 6; #145 breaks it and looks flaming; pair P14 with #150, which is labelled Flaming and looks silver); every fifth position from 4 is a Cyber cryptoGAN (12 of 12; #165 and #185 are off the rhythm); the six positions with a single 1 in binary are all special (#130, #133, #161 colour flash; #131, #137, #145 the only cyberGANs, with eyes drawn as binary digits). Dimensions 8 marks four pieces (#129, #130, #141, #161), all with still eyes and the colour flash; #174 has both traits but Dimensions 4. The Eyes trait only ever says 0, 2, 4 or 8, though fourteen robos show three eyes. Essences continue the wordlist run. Pindar stopped using the name formula: most names are compounds (Rainbot, G1itch1ng, B1-b0rg). Four skull/robo hybrids (#161, #175, #187, #191).

### Drop 3, spookies (#193 to #208, 17 October 2021)

Sixteen pieces counted in octal, 0o0 to 0o17. The hidden-type letter follows the octal count; the names follow the plain count from 1: the four Ghostly spookyGANs sit on the squares 1, 4, 9, 16 (chance 1 in 1,820). Dimensions 5 only on Fibonacci positions, as in the skull drop. Every piece wears a costume except #206 Skullbot, which looks like a robo-drop hybrid. Six consecutive pieces (#197 to #202) hide Ghost-drop features under the costume, though the Ghost drop came later. #193 at position zero carries the hidden-type letter F, as does robo #129 at position zero. #207 looks like a Review-drop piece and carries a painted "4".

### Drop 4, Solstice (#209 to #300, 21 December 2021)

A calendar. Sol (#209, "Day") opens it, Luna (#300, "Night") closes it, and between them are five rows of 18. Rows 2 to 5 open with a tribute to another artist dated on an equinox or solstice (#228 March 20, #246 June 21, #264 September 22, #282 December 21). Every other piece is one frame of one of eight season videos, named in the Arweave filename with a frame number. The background pattern is the video's signature: blue five-point snowflakes (winter), green four-dot marks (spring), yellow squares (summer), orange dots (autumn), and a mix for a transition. Inverted pieces show the pattern in negative. Each row is one year and the dates never run backwards inside a row. Frames arrive in threes from one video; each video's frame numbers only increase down the drop. Four videos supply exactly eleven frames each (winter, spring, spring-to-summer, summer-to-autumn). Essences run the wordlist in order (index 128 to 223), mint order equals piece order. Breaks: row 2 (#229 to #245) starts at the equinox, runs only March to August, has no autumn, alone uses the "update" video (batch code 44 where the others are 04 to 10), and holds the only out-of-order frames (#240 to #242, essences before, begin, behave). #240 and #297 have dates against their filenames; the picture sides with the filename. #259 and #268 are the two out-of-sequence essences. #298 is labelled Twin and is not one; #281 is labelled Glitch and does not glitch by eye.

### Drop 5, Ghosts in the Machine (#301 to #444, 31 January 2022)

A colour wheel run nine times. Hue advances 22.5 degrees with every piece, so every 16 pieces run the rainbow once (measured from the frames: average miss 17 degrees; random would be 90). Each of the 16 colour slots holds, over the nine cycles, 4 ghostGANs, 2 testGANs, 1 spaceGAN, 1 monsterGAN and 1 skullGAN, so spaceGAN, monsterGAN and skullGAN each run the rainbow exactly once (16 each), testGAN twice (32), ghostGAN four times (64, less the two one-offs #327 Dry iceGAN and #407 Mint Bot, which take two ghostGAN places in the cyan slot). The colour-word SubTypes follow the slots: Bloody in magenta-to-red, Glowing in the reds, Shy in the yellows, Vivid in the greens, Frozen from green to azure, Ultra from azure to violet. The other SubTypes carry no colour and no shared look. Each species keeps one look (testGAN: test-pattern background, robo eyes, TEST for a mouth; spaceGAN: starfield, five-bar eyes, invader badge; ghostGAN: empty white square eyes on binary; monsterGAN: pod-symbol background, many eyes). Essences are not a wordlist run here; the mint order reverses three times, all in cyan slots. Breaks: #366 is the one piece far off its slot colour; twelve colour-word labels sit in a slot of another colour (#309, #313, #357, #359, #361, #371, #377, #380, #393, #403, #429, #430), and on all but #357 the picture follows the slot, so the word is the odd part; #306 has spaceGAN eyes on a ghostGAN; Ultra and Vivid supply exactly eleven pieces each.

### Drop 6, Act 1 Review (#445 to #512, 23 August 2022; four tributes shown at the clue in February 2022)

Four tributes open it, #512 Quantum glitchGAN (essence 'final', Dimensions 11, the only xxxxxGAN) closes it. Pindar's filenames sort the rest into working groups that walk back through the earlier drops in order: reprise00 to 03 are pairs of skulls, 04 and 05 robots, 06 cryptoGANs, 08 spookies, then Solstice and Ghost (09 to 12 with the testGAN and monster groups), then new groups: skullall, reprisehighdef, repriseaa (four derivatives of Mint Bot), reprisebs (Inverted skulls), reprise13 (the seven abstract pieces the collector calls Review Drop Natives). The counters keep counting the old drops: #457 to #464 are 64 to 71 in binary and #497 to #500 are 72 to 75, so the robo count runs 0 to 75 across two drops; the spooky reprises are 0x20 and 0x21 (32, 33) in hex. Groups come in pairs and fours. Colour runs warm to cool and does not cycle. Twelve Review essences form a run of the wordlist from 'solution' (index 1655) to 'south' (1666): #450 solution, #453 solve, #458 soup, #463 sorry, #471 someone, #479 song, #484 soon, #487 soul, #494 sort, #499 sound, #502 source, #508 south. Breaks: no reprise07; the test group is split (#467 to #469 saved months early, at the wrong speed, with a new scrambled TEST background; #481 to #484 later, traditional); reprise11 numbered ahead of 09 and 10, and #477 filed as 11 among them; Inverted pieces sit on black except #509 and #511 on white; #447 minted out of order; #449 a Shimmer that does not shimmer; #456 a Ghost that does not fade; #508 and #510 labelled Bot with skull or ghost eyes; #477 and #478 carry the Solstice spring-to-summer background; #479 and #480 hold a frame and carry a new beige background.

## 5. Facts that cross the drops

- **Sizes.** 128 + 64 + 16 + 92 + 144 + 68 = 512. The first three are 2^7, 2^6, 2^4. The Ghost drop is 9 × 16. 512 = 2^9 = 64 bytes.
- **Counters.** Decimal (skull), 6-bit binary (robo), octal (spooky), printed number (Solstice), n/144 (Ghost), n/512 (Review), with the robo binary count continued to 75 and the spooky count continued in hex in the Review.
- **A BIP-39 word is an 11-bit index** into a 2,048-word list. Twelve words are 132 bits: 128 bits of entropy plus a 4-bit checksum. "Eleven glitches" and "11 bits per word" is a coincidence worth a first-principles look (card 150, "Read the eyes as ones and zeros", is open). Only one attempt has touched it (a 32-by-32 pixel-difference grid read as 11-bit indices, swept with no hit, D-notes line 687). One related fact: the chain description of #130 qubitGAN says "1024 animations" where every other robo says 2048, and 1024 is half the wordlist.
- **The prime rule** holds in the skull and robo drops by eye and by file: a piece glitches when its position is a prime that is not Fibonacci. Seven skull primes carry another Type and still glitch. In the Solstice drop, five of eight Glitch-Type pieces sit on primes counting from 1 (weak).
- **Fibonacci positions carry 5 Dimensions** in the skull and spooky drops. Across the first three drops only #7 has 5 Dimensions off a Fibonacci position.
- **Pindar's index column.** His January 2022 working table gives a BIP-39 index beside every piece from 1 to 300; the essences of drops 1 to 4 run the wordlist in order, with displacements (#21, #57, #259, #268). From #301 on the essences are scattered.
- **The eleven pieces that play at the wrong speed**, across three drops, holding both 'random' (#468) and 'final' (#512): #155, #176, #180, #182, #190, #228, #446, #467, #468, #469, #512. Run as a set (card 237), no match.
- **Labels the picture does not support, by eye, across five drops** (Glitch word aside): eleven pieces (#57, #105, #156, #169, #456 Ghosts that never fade; #143, #168, #449 Shimmers that do not; #298 Twin; #381, #423 colour words). Run (card 244), no match.
- **Deliberate pairs** P01 to P15 (pairs.json), including #14/#75 byte-identical files, #479/#480 held frames, #209/#300 day and night, #363/#473 TRUE and FALSE, #411/#450 key and solution, #145/#150 swapped Silver/Flaming, #7/#8 swapped Dimensions.
- **Same names.** 91 names are shared by two or more pieces (the site's Same names page, /names/).
- **The hidden-type letter** on chain encodes the number class for drops 1 to 3; position zero of the robo and spooky drops carries F, outside the scheme.
- **Colour words** in essences: the ten plain colour words are each an essence exactly once (plus rose and bronze).

## 6. How to build a hypothesis

The collector's instructions on creativity and construction:

- **Every hypothesis has three parts, and each part must be stated.** (1) How the eleven glitches are selected. (2) How each selected piece becomes a word. (3) How the eleven words are ordered. A proposal that names only the pieces is not a hypothesis yet.
- **Each part may have several answers.** Selection, word rule and ordering are separate choices, and Pindar may have used a different answer for each. Do not assume the word is always the essence, or that the order is always a sorted column. Enumerate the plausible answers for each part and run the combinations.
- **The solution may take more than one step.** A first step may produce an intermediate result (a set of pieces, a set of numbers, a pointer to other pieces) that a second step turns into words or an order. Chains of steps are allowed and expected.
- **The steps may differ by drop.** Because every drop has its own logic (section 4), the rule that selects or orders a piece from the skull drop may not be the rule for a piece from the Ghost drop. A hypothesis may say: "in the drops counted in binary, use the binary position; in the Solstice drop, use the calendar date; in the Ghost drop, use the colour slot." State the rule per drop and why the drop's own logic points to it.
- **Be creative, then be exact.** Generate freely across the lenses in section 7, but every proposal that goes to a run must be reduced to the three parts above in one or two plain sentences each, so that a sweep can be built from it and its result recorded against exactly that statement.

## 7. Suggested shape for the swarm

Give each agent one lens and one deliverable: a list of statable hypotheses (which pieces, which word rule, which order, and why the collection's logic points there), each checked against the registry before it is proposed. Lenses the collector asked for (each proposal reduced to the three parts in section 6):

1. **Bits and bytes.** How would a programmer organise 512 items, six drops of these sizes, 16-slot cycles, 6-bit and octal counters, 11-bit words? Which pieces sit at the binary landmarks (powers of two, palindromes, one-hot numbers, bit-reversals, complements) across the whole 1-to-512 count, not just within a drop?
2. **Breaks.** Every piece that breaks its own drop's logic, gathered across drops (drop_rules.json lists them). Which breaks are Pindar's hand and which are noise? Which elevens do they form, and with which pieces left out?
3. **Echoes.** Logic from one drop reappearing in another: the Fibonacci-5-Dimensions rule in skulls and spookies; F at position zero in robots and spookies; Ghost-drop features under spooky costumes; Solstice backgrounds on Review pieces; the robo binary count continuing in the Review; #207 and #407 looking like the Review before it existed.
4. **Each drop as a map to the Review.** For every Review piece, what earlier piece or rule does it reprise, and what did Pindar change in the reprise? The reprise map (review_reprise_map.json) and the collector's Solution Block Review (D87) are the starting points. The change may be the glitch.
5. **The Review as a map back.** Read the Review's own order, groups, counters and the 'solution' block as instructions pointing at earlier pieces.
6. **Visual logic.** Only the collector's eye counts (visual_audit_2026-09-27c.json, 300-plus notes; his SubType descriptions; the by-eye tags). Which visual classes are exact, small and cross-drop?
7. **Numeric logic.** Number classes, positions, counters, mint numbers, frame counts, dates, and the wordlist indices of the essences, treated as one system.
8. **Adversary.** For every candidate the others produce, argue why Pindar would not have built it, and check it against the three tests.

Deduplicate against the registry (`theories/_registry.json`, keyed by the sorted eleven words) and the tested-set files on the site before running. Report to the collector in plain sentences: what each theory is, why the agent thought it valid, what was run, and what happened.

## 8. Open hypotheses to carry forward

Cards marked open on the site: 239 and 240 (the number-class look rules and their eleven breaks: #7, #8, #16, #31, #41, #55, #89, #145, #165, #174, #185, to be revisited now that every drop is audited); 150 (eyes as ones and zeros); 171 (structural glitches encode the word order, not the words); 178 (substitutions echo what you can see); 179 (two candidate elevens with the word registers closed); 57, 78, 80, 91, 123, 140. The So-Word analogues page (choosing drop-1-to-5 analogues for the eleven 'so' words) awaits the collector's picks.

## 9. Data map

### Public site, https://bitganspuzzlebook.com

- `/` the collection with every tag; a drop tag opens with its layout and rules. Share links: `/?tag=drop%3Aghost` and any tag id.
- `/names/` the same-names page.
- `/brief/` the earlier standing brief for outside models (still valid background).
- `/data/` 131 files in tiers. Primary sources: `our_spreadsheet.csv` (68 columns, Pindar's sheet plus our columns), `arweave_bitgan_properties.csv`, `arweave_file_txids.json`, `arweave_filenames.json`, `chain_raw_by_mintindex.json`, `all_onchain.json`, `transfers.json`, `bip39.txt`, `page015_jan1/bip39_index.json` (Pindar's index column), the journal and news texts under `journal/`. Derived: `drop_rules.json` (every drop's layout, rules and breaks), `DROP-NUMBER-RULES.md`, `NAMING-CONVENTIONS.md`, `ARWEAVE-LAYER.md`, `primary_facts.py`, `gif_facts.json`, `save_sessions.json`, `first_transfers.json`, `review_reprise_map.json`, `solution_key.json`, `global_numclass.json`. Reading: `anomaly_ledger.json` (every anomaly per piece, additive), `visual_audit_2026-09-27c.json` (the collector's complete audit), `michael_visual_notes.json`, `pairs.json`, `taxonomy.json`, `glitch_dictionary.json`, `DECISIONS.md` (D1 to D110), the swarm of 21 September under `swarm_2026-09-21/`, `curated_elevens.md`, `SOLUTION_SPACE.md`, `METHODS_RECIPES.md`, `REGISTRY_POLICY.md`, and the coverage files `tested_sets_no_hit*.txt` (about 290,000 lines) and `coverage.json`.
- API, `https://api.bitganspuzzlebook.com`: `/api/pieces`, `/api/tags`, `/api/hypotheses` (244 cards), `/api/runs`, `/api/records`, `/api/concepts`, `/api/principles`, `/api/search?q=`, `/api/stats`, `/api/export.csv`; media at `/media/gifs/N.gif` and `/media/thumbs/N.png`.

### Local repository, `~/Desktop/bitgan_puzzle_work`

- `primary_facts.py`: SHEET, CNAME, TI (mint slot), SAVE_TIME, FIRST_TRANSFER, FILENAME, REVIEW_GROUP, DIFFS, WRONG_SPEED, numclass, ctraits, bip. Import it; do not re-derive.
- `tag_schema.py` and `bitganspuzzlebook/scripts/build_tags.py`: every tag with its definition and members. `drop_rules.json`: the per-drop logic shown on the site.
- `theories/DECISIONS.md` (D1 to D110), `theories/anomaly_ledger.json`, `theories/_registry.json`, `pairs.json`, `lore_catalog_runs/` (visual audits, ghost colour cycle in `ghost_subtype_review/colour_cycle.json`, Solstice calendar, same names, so-word analogues, swarm files), `workbench/gifs/` (the 512 original GIFs), `sweeps/` (every past run with its task list and logs), `build_hyp62.py` to `build_hyp95.py` (patterns for building runs).
- Harness: `sweeplib.py`, `qbuild.py`, `qrun2.py`; recording: `bitganspuzzlebook/scripts/push_findings.py`, `update_all.sh`.

### What "everything we have" does not include

We do not hold the 2022 CryptoVoxels galleries Pindar referred to. We hold no private message from Pindar. Everything else he published (journal, news posts, sheet, Arweave files, chain) is in the data bundle.

## Corrections (28 September 2026)

These corrections are added after the second swarm attempt. The text above is left as it was written.

- **Robo essences (Drop 2 above).** The brief says the robo essences continue the wordlist run. That is wrong. Pindar's own index column shows the robo essences are chosen by hand and scattered from word 57 to word 2046. They do drift up the alphabet through the drop: the rank correlation with position is 0.75 over positions 0 to 55 and 0.48 over all 64. The last row, #185 to #192 (digital, endless, mirror, pattern, blue, primary, error, zero), breaks the drift. Those eight were saved with #184 in their own session on 27 September 2021, the night before the roboGAN puzzle.
- **Fibonacci positions and 5 Dimensions (Drop 3 and the cross-drop facts above).** The brief says Fibonacci positions carry 5 Dimensions in the skull and spooky drops, and lists this as an echo between the two drops. In the spooky drop the rule holds one way only. Every spooky piece with 5 Dimensions (#194, #198, #206) sits on a Fibonacci position, but the other four Fibonacci positions (0, 2, 3 and 8, counting from 0) have 3 or 4 Dimensions. So it is not the same rule as in the skull drop, and the brief oversold it (see also D112).