# Brainstorm synthesis: using 32 to decode the Act 1 puzzle Seven lens-agents (arithmetic, grid, structure, binary, pointer, ordering, intersect) each proposed ideas for using the number 32 to find or read the eleven glitch words. This document merges their ~50 raw ideas into a deduped, ranked list. ## Dedup note Several ideas showed up in three or more lenses wearing different names. I merged these into single entries and credit every lens that found them: - **The ±32 "hop" idea** — add or subtract 32 from a piece's number to find a partner piece — was proposed independently by arithmetic (as an index offset), binary (as an XOR-32 bit-flip, which is the same move for these numbers), structure (as a stride), pointer (as a "transplant operator," with an actual test run), and intersect (same test, different framing). All five point at the same mechanism. Merged into one entry. - **The 32x32 pixel-grid, glitch-frame-diff idea** — find the most-changed cell in a piece's own artwork and turn its (row, column) into an 11-bit wordlist index — was proposed by both the grid lens and the structure lens with nearly identical math. Merged into one entry. - **"Cut the collection into sixteen blocks of 32" idea** — group pieces by mint-position ÷ 32 and check which block each ratified class falls in — was proposed by structure, grid, and intersect. Merged into one entry. - **The robo-drop "position P pairs with position P+32" idea** — appeared as its own structural idea, as the binary XOR-32 idea, and as one case of the general ±32 hop. Folded into the ±32 hop entry as its cleanest, already-confirmed special case. - **"2048 = 64 pages of 32 words" idea** — splitting the bip39 wordlist itself into 32-word pages — was proposed by both arithmetic and structure. Merged. - **"bitGANg 1001's own curated 32 pieces" idea** — proposed by both structure and intersect. Merged. ## Cross-check against the dead list and DECISIONS A few ideas turned out to already be closed, or to quietly re-run something closed: - **GAN-as-bip39-index reproducing a piece's own Essence word** (arithmetic idea 1) — the pointer lens checked this directly and found it is NOT special to piece #32: the same match (Essence = bip39[GAN-1]) holds for the entire skull drop (1–128) except two gaps. Pindar simply flavor-texted the skull essences from the wordlist by design. For skull pieces, this idea is just the already-DEAD Essence register wearing a new name. It stays live only for non-skull pieces (robo/spooky/etc.), where that baseline pattern hasn't been shown to hold. - **Position-32-of-each-drop probe** — explicitly DEAD per the brief. Two lenses (arithmetic, intersect) proposed the *collection-wide* mint-modulus-32 version instead of the per-drop version. That's a different, not-yet-dead test, but arithmetic already ran it locally: it produces 17 members, not 11 — a clean miss, not a hit. Demoted to low priority. - **HODL +128-partner** — DEAD per the brief. The ±32 hop (above) uses a different constant and a different mechanism (structural bit-flip, not flat addition), so it is not a re-test — but it's close enough in spirit that both the structure lens and the pointer lens flagged the overlap honestly. - **Montage-as-selection** — DEAD per the brief, but that refers to the official 512-piece display montage. bitGANg 1001 is a different, real, Pindar-curated object (a 32-piece auction promo). Nobody has ever pulled its actual 32-piece roster and checked it against the glitch registry — worth one cheap look before assuming it's the same dead idea. - **32nd glitch in mint order**, **mint-index%32==0 filter**, **mint-rank bracket [97,243] as a class window**, and **#32/#161 as adjacency slot-markers** were all already run by the pointer lens with negative or weak results. Kept in the record but demoted to Tier C. - **Eyes as a literal 5-bit register** — disproven on the actual data (eye counts only ever run 0–10, nowhere near 32). This closes the literal version of the still-open H-2 "eyes-as-binary" hypothesis from D27/D28 without resolving H-2 itself. ## Tier A — cheap, genuinely new, plausible: test these now 1. **±32 hop across ratified-class piece numbers.** Add or subtract 32 from a piece's catalog number and see if you land on another already-flagged glitch piece. This is NOT the dead +128 idea — it's a different number and (for robo pieces) it's actually flipping a single binary digit, not just adding. Two lenses already ran quick versions and found real hits: piece #129 shifted by 32 lands exactly on #161 (both already in the Still-Eyes class), and a handful of Fibonacci-eleven and Palindrome members shift onto other ratified-class members at a noticeably higher rate (about 46%) than you'd expect by chance. **Test:** write a short script that shifts every member of every ratified class (Twin, Entangled, Palindrome, Still-Eyes, Backwards- Fib-Eleven, HiddenType-misfiles) by +32 and -32, log every landing that hits another flagged piece, and compare the hit rate against a random-shift control (try shifts of 16, 48, 64 as well) to make sure 32 is really special and not just "the drop is dense." Promising because three independent lenses converged on it without knowing about each other. 2. **Glitch-frame pixel diff → grid coordinate → wordlist index.** Every artwork is a 32x32 pixel grid. For each of the eleven candidate pieces, find the single grid cell that changes the most between its glitch frame and a clean frame. That cell's row (0–31) and column (0–31) give you 10 bits; one more bit for "brighter or darker" gives exactly 11 bits — the precise size needed to point at one of the 2048 wordlist words, with nothing left over. **Test:** for the Backwards-Fibonacci-Eleven class (the class already defined by a frame-level glitch event), load each piece's GIF, diff consecutive frames, find the most-changed 32x32 cell, compute the word index, look it up in bip39.txt, and run the resulting 11-word set through the existing sweep harness. Genuinely new — nothing has read actual pixel data before; all prior work stayed inside the metadata fields. 3. **The doubling eleven: {0,1,2,4,8,16,32,64,128,256,512}.** Pindar's own published HODL list is eleven Fibonacci numbers built by adding. This is the same shape built by doubling instead — ten powers of two plus #0 as the anchor, and 32 is the sixth term. It's the direct multiplication counterpart to a list the puzzle already treats as meaningful. **Test:** pull each piece's essence and mint-order position from deck_512.json, assemble the eleven words in a few orderings (mint order, ascending value, #0-first, #512-last), and run all variants through the sweep harness. Cheap — the set is already fully defined, no new data needed. 4. **{#32, #95} as the missing 10th pair.** D2 still has an open slot for a 10th deliberate pair to complete the pairs-plus-anchor eleven. The chain burn log shows #32 and #95 were burned in the exact same transaction — the same same-timestamp signal that already found other confirmed pairs in this project. Nobody has logged this pairing before. **Test:** confirm the shared burn timestamp (quick lookup), check #95 for any independent visual or trait link to #32, and if it holds up, substitute {32,95} for the current placeholder candidates in the pairs+anchor sweep. 5. **bitGANg 1001's actual 32-piece roster.** This is a real, different object from the dead "montage-as-selection" — a 32-piece promotional GIF Pindar hand-picked for a pre-clue auction, already saved locally as frame images. Nobody has ever identified which 32 pieces are actually in it. **Test:** match the extracted tiles against deck_512.json to get the 32 GAN numbers and their on-screen order, then intersect that list against every ratified glitch class. If the overlap lands at exactly 11, that's a strong new selection with a built-in reading order. 6. **Mint-index as a direct bip39 pointer (non-skull pieces only).** For each piece, look up bip39.txt at the position given by its own place in the overall mint sequence (not its essence, not its catalog number). This is only worth running on robo, spooky, solstice, and ghost pieces — the skull-drop version of this idea turned out to just re-derive the already-dead essence register (see cross-check above), so skull pieces should be skipped. **Test:** for the non-skull members of the Backwards-Fibonacci-Eleven and Prime-Position-Robos classes, compute bip39[mint_index - 1], assemble in mint order, run through the sweep harness. 7. **Sixteen blocks of 32: does any class put exactly one member per block?** Cut the 512-piece mint sequence into 16 consecutive chunks of 32 pieces. Count how many members of each ratified class fall into each chunk. If a class puts exactly one member in eleven of the sixteen chunks, that's a strong "which chunk, in what order" signal that needs no manual guessing. **Test:** this was partially run already (two classes checked, neither fit) — extend the same histogram check to Twin, Entangled, Still-Eyes, Pairs, and Prime-Position-Robos, which haven't been checked yet. ## Tier B — worth building, but slower or less certain - **Base-32 digit-pair radix reconstruction.** Convert a piece's catalog number or mint number into two base-32 digits (since every relevant number is under 1024), then recombine those two digits a few different ways into an 11-bit wordlist pointer. Test: write the conversion, try all four simple recombinations, run each through the sweep harness. - **Wordlist read as 64 pages of 32 words.** Split bip39.txt itself into 64 chunks of 32 words each. Use a piece-native number as "which page" and another as "which word on the page," instead of one big index. Test: try GAN-number-mod-64 for the page and mint-index-mod-32 for the in-page slot, look up the word, sweep. - **New reading orders on the 32x32 structure.** Several cheap ways to re-order an already-defined eleven before reading the words off: sort by (piece number mod 32), read the grid in a snake pattern (left-to-right, then right-to-left, alternating every 32), reverse the low 5 bits and high 4 bits of each piece's position (a bit-reversal shuffle), or start reading from wherever piece #32 itself sits in the mint sequence and walk around the loop from there. Test: each is a few lines of code — re-sort an existing candidate eleven by the new key, assemble the words, run through the sweep harness. Caveat: the essence register is already extensively dead across many orderings, so these are worth trying mainly if paired with a register that hasn't been swept yet, not essence alone. - **32-row binarized bitstring sweep.** Shrink each piece's artwork to a true 32x32 grid, turn one row into a string of 32 light/dark bits, convert to a number, and take that number mod 2048 as a wordlist index. Try all 32 possible rows (cheap — only 32 trials) and see if any single row produces clean, distinct English words across all eleven candidate pieces. - **Direct pixel-coordinate read (no diff needed).** Simpler cousin of the Tier-A pixel idea: instead of finding the most-changed cell, just use the piece's own catalog number or drop-position directly as a (row, column) address on its own 32x32 grid, and read whatever color sits there. Faster to test than the frame-diff version, but less principled. - **Distance-from-32 on the Naga/Still-Eyes classes.** #161's binary position is exactly 32; measure how far each of the 8 Naga-palindrome and 8 Still-Eyes pieces sits from that number and use the distances as a fresh non-essence register. - **Base32-text-alphabet decode.** Standard Base32 has 32 letters/digits (A–Z, 2–7). Map each piece's number-mod-32 through that alphabet and see if the resulting short string spells anything recognizable (KEY, SEED, etc.) — a plain-text hint rather than a wordlist pointer. ## Tier C — long-shot or already mostly tested negative - **Literal "go to cell 32" self-read on #32 and #161's own art.** Interesting as a tie- breaker for the Tier-A pixel idea, but only two data points — not a standalone test. - **Twin split-grid pixel read to prune Twin from 13 down to 11.** Conflicts with Michael's own ruling (D25) that Twin is context, not a glitch source right now — low priority unless that ruling changes. - **Mint-rank window [97, 243] as a class-validator.** Already tested: it's too coarse (30+ pieces fall inside it), only weakly picks out Palindrome/Still-Eyes. - **32nd glitch piece in mint order.** Already tested: lands on GAN #126, not a member of any ratified class. Null result. - **Collection-wide mint-index-mod-32==0 filter.** Already tested: gives 17 pieces, not 11. Clean miss. - **#32/#161 as adjacency "which slot needs a hop" markers.** Already tested: only weak, chance-level adjacency to existing classes. - **Eyes-as-5-bit register / eye-count-sums-to-32 checksum.** The literal per-piece version is disproven (eye counts never get near 32). The weak "does the eye-count sum across a whole class equal 32" version has been tried on two classes and missed by 2–3 both times. - **Word-list-index-mod-32 as a reading order.** Sorts already-chosen words by their own wordlist position instead of by piece order — cute but the lens itself rated this low confidence. ## Strongest intersections with ratified classes - **±32 hop ↔ Still-Eyes class**: #129 and #161 are both already Still-Eyes members, and they differ by exactly 32 (and by exactly one binary digit). This is the cleanest, most structurally explained hit in the whole brainstorm — worth leading with. - **±32 hop ↔ Backwards-Fibonacci-Eleven / Palindrome / Still-Eyes**: several other ratified-class members land on each other under the same ±32 shift (148→180, 160→192, 190→222), suggesting 32 may be a real structural link between these classes, not just noise. - **{#32,#95} burn pair ↔ D2's open pairs slot**: directly answers a pending decision rather than opening a new one. - **bitGANg 1001 ↔ every ratified class at once**: if the 32-piece roster overlaps a class at exactly 11, it would settle both selection and order in one shot. - **32x32 pixel grid ↔ Backwards-Fibonacci-Eleven**: this class is already defined by a frame-level glitch event, making it the natural first test case for the pixel-diff idea — the mechanism and the class are already aligned. ## Top 5 to run first 1. **±32 / bit-flip hop across all ratified classes** — script the shift, compare hit rate against random-shift controls, focus on the confirmed 129↔161 Still-Eyes link. 2. **Pixel-diff → 32x32 grid cell → 11-bit wordlist index**, run on the Backwards- Fibonacci-Eleven pieces first since that class is already glitch-frame-defined. 3. **The doubling eleven** {0,1,2,4,8,16,32,64,128,256,512} — assemble, sweep a few orderings, cheap and self-contained. 4. **{#32,#95} as the 10th pair** — confirm the shared burn timestamp, then plug into D2's pending pairs-plus-anchor sweep. 5. **Identify bitGANg 1001's real 32-piece roster** and intersect it against every ratified class, checking for a clean 11-piece overlap. --- ## FIRST RESULTS (run 2026-08-31, same session) - **Top-pick #1 ±32 hop — KILLED by control.** Actual +32 landing rate across all flagged pieces = 20.0%, LOWER than shift-16 (30.6%) and no better than shift-7 (19.2%) or shift-11 (21.9%). 32 is not a special hop distance; the landings are an artifact of the robo drop being densely flagged at binary-position intervals. The "46%" estimate was wrong. The specific 129<->161 link (both still-eyes, differ by the 32-bit) survives as a structural note, not a mechanism. - **Top-pick #3 doubling-eleven {0,1,2,4,8,16,32,64,128,256,512} — SWEPT DEAD.** Essence (universe abandon ability about abstract acid advance among avocado bike final) + piece# + mint registers x 4 orderings, plant-validated, no hit. Note: the powers-of-2 skulls carry bip39[2^k-1] as essence (the by-design skull pattern). - **Top-pick #4 {#32,#95} burn pair — VERIFIED.** #32 (Slender skullGAN, advance) and #95 (Flaming Ghost skullGAN, armed) burned in the SAME transaction 2021-10-14T14:49:06. Logged as a candidate for D2's open 10th-pair slot. (Other same-tx burn groups: the 4 spookies 0o14-0o17; the 4 skulls 74/49/20/11.) - STILL OPEN (need image work): #2 pixel-diff grid-cell -> 11-bit index on the fib-eleven; #5 bitGANg 1001's 32-piece roster identification + class intersection. ## MORE RESULTS (2026-08-31) - **Top-pick #2 pixel-diff 32x32 grid-cell -> 11-bit index — TESTED DEAD.** The glitch on the fib-eleven is a re-inserted WHOLE animation frame, so the whole figure redraws and the "most-changed cell" is diffuse (top1/top2 ratio 1.00-1.07, no spike). 16 variant word-sets (prev/frame0 ref x row/col-major x 0/1-based x maxcell/centroid) swept, plant-validated, no hit. The pixel-cell-argmax reading does not encode the message. Forensics + sweep agree. - **Top-pick #5 bitGANg 1001 roster — DEAD END (premise fails).** bitGANg 1001 = a single Verisart-certified NFT (22-for-2022 auction) whose art is "an animated GIF of 32 unique bitGANs" = rainbow_ghosts_10x6_0_920.gif (an 8x4=32 interior of rainbow ghost faces inside a binary-code border). Two problems: (1) shape-matching the 32 tiles to the 512 deck gave only weak/ambiguous scores — no reliable GAN mapping; (2) more fundamentally, the lore frames these rainbow ghosts as a "sneak peak" at the NEXT drop's color style — a distinct generation, NOT members of the 512 deck the glitch classes index. So the 32-vs-classes intersection is not even well-defined. No fabricated roster. Closed. ## TIER-A SCOREBOARD (final) ±32 hop = dead (control). doubling-eleven = swept dead. pixel-grid = swept dead. bitGANg 1001 = dead end (premise). {#32,#95} burn pair = VERIFIED, pairs-sweep pending. => 4 of 5 top picks resolved negative; the one live residue is the {#32,#95} pair for a pairs+anchor sweep.