# the collector's standing decisions log ## D1 — Spooky drop glitch (2026-08-27): ALL THREE as variants the collector's ruling: either #205 or #207 could be the glitch, PLUS a third candidate he added — **#206 "Skullbot spookyGAN" (0o15), essence `grain`** — the only spooky NOT in a Halloween costume (it is a robot; the drop's theme is costumes, so being uncostumed breaks the theme). Selection among {205 grace D4, 206 grain D5, 207 grant D3} depends on the selection criterion in play; be flexible and RUN ALL THREE AS VARIANTS in candidate solutions. The three cover Dims {3,4,5}, so Sudoku assemblies can choose per slot need. ## Pending - D2: ratify the ten deliberate pairs + which anchor completes them to eleven. - D3: Sudoku backbone review. - D4: true-class namings for the mislabel eleven. ## D2 — Pair ratification (2026-08-27): 9 confirmed + 1 rejected + replacement Grades: CONFIRM 14+75, 479+480 ("nothing else in the collection with these colors"), 209+300, 489+491, 130+133, 165+207 (confirmed but "suss — could be coincidence of the inference"), 363+473 (+ the collector's open idea: TRUE/FALSE may be a statement on each piece's relationship to its own drop), 175+205 (+ #175 is a skull/robot hybrid, unusual for its drop in other ways), 16+100 (the collector conditional-confirmed; byte-verified: both chain names carry "Primoridal", the only two in 512 -> CONFIRMED). REJECT 150+179 ("179 is definitely a ghost" — label fits, no swap; 150 borderline). Candidate replacement 10th events: {465,466} hex-radix pair (0x20/0x21, name-visible, mint-adjacent 607/608 — note mint indices EXIST, older mint=None notes are stale) or {390} double-mint. ANCHOR: use #357 seed OR #512 final as the 11th, never both; if seed, also try phrase- INITIAL position (the collector's instruction). Sweeps: q0827d — all 12 essence-register variants (hex/dbl390 x first/last-speaks x 357mint/357first/512last) NEGATIVE, plant-validated. The pairs SELECTION stands; the essence register on it is now exhausted -> next registers: pair-relation reads. ## D3 (partial) — Sudoku review (2026-08-27): brute-force the closed space the collector: "everything you showed me seemed viable" -> run the FULL pairs-x-Sudoku closed space (127,920 orderings). Weak point flagged: "the first slot of the shimmerGANs seemed weakest" = slot 2 / Dim-2, occupied by the Shimmer splice twins {67,71}. If the space dies, slot 2 is the first suspect: hunt alternative Dim-2 candidates beyond the shimmers. ## D5 — Burnt-piece eligibility (2026-08-27): RUN BOTH AS VARIANTS Finding (the collector spotted): all 32 burns happened BEFORE the clue (Oct12-Dec17 2021; clue Feb 5 2022). Burns only in first 3 drops: 16 skull, 6 robo, 7 spooky; none in solstice/ghost/review. Burnt GANs: see burnt_gans_set.json. ALL THREE D1 spooky candidates (#205,#206,#207) are burnt; #204 in the mislabel-11 is burnt. In pairs/sudoku pools only #205,#207 are burnt. RULING: burnt-status is an AXIS, not an exclusion. Sweep candidate sets BOTH with and without burnt pieces. Non-burnt spooky survivors (the 9 real-at-clue-time spookies): #193,195,196,197,199,200,201,202,203. Deck now carries a `Burnt` boolean. GUARDRAIL for SOLUTION_SPACE: add "burnt-eligible vs burnt-excluded" as a selection axis. ## D4 — Mislabel true-class review (2026-08-27): CLASS MOSTLY FAILS + 2 real mislabels the collector graded 11 mislabel candidates + 3 blind controls. Result: the mislabel-eleven as a SELECTION does NOT hold — 9 of 11 labels actually MATCH the art to his eye. The "11 label-liars" glitch class is CLOSED as a selection (independent of word/order — this is a triple-discipline SELECTION kill, not a word/order kill). CONTROLS validated his eye: #300, #179, #67 correctly marked "matches"; and he caught #357 (which I slipped in as a control) as a genuine contradiction. Two REAL mislabels surfaced: - #423 Rainbow skullGAN — "almost entirely blue and greens, not rainbow" (ess cruel, D5). - #357 Shy skullGAN (OUR SEED ANCHOR) — "doesn't hide/disappear at all, don't know why it'd be Shy" (ess seed, D4). Its Shy SubType is suspect; #357 already over-determined as a color-glitch (yellow-where-cyan). True class not yet named by the collector. UNSURE: #204 (clown not zombie, but BURNT flaming piece used to forge Cyber Zombie 10,366; the flaming-forge requirement explains its Flaming SubType), #150 (has red=flaming but a long fully-blue stretch — possible contradiction; already flagged 'Art contradicts SubType' + 'No on-chain traits'), #511 (not that inverted, "common in corrupted GANs"). DATA UPDATE: #491 Frozen skullGAN — the collector confirms it has the COLOR-SPEEDUP effect (log it; it's the scanline-pair member 489+491). #204 has 'Burnt in forge' flag confirmed. ## D4b — True-class descriptions for the two real mislabels (2026-08-27) #357 (anchor): skullGAN, mostly YELLOW top / blue bottom; ALMOST an invader icon on the shirt but not quite; head shape reminiscent of Sol/Luna but the left/right/top spikes never quite materialize. => a piece that ALMOST becomes several other things and doesn't. True-class BIP39 candidates: yellow (over-determined — matches his original "yellow where cyan expected" color-glitch finding), blue, half, almost, shy, alien, spike. #423: blue/green skullGAN with skulls in background; brief pop of color "like the RECYCLING symbol of the Eclectic Method artist GAN". => references Eclectic Method (one of the 4 original cryptoArtGANs). True-class BIP39 candidates: recycle (standout — the Eclectic Method motif), blue, green, echo, mirror. PATTERN NOTICED (candidate NEW glitch class): both confirmed mislabels are INCOMPLETE / NEAR-MISS HOMAGES — the piece almost renders as another entity (invader, Sol/Luna, Eclectic Method) but fails to complete. Fits Pindar lore ("trying to come into focus but couldn't"; "transforming into a different bitGАН"). Worth a census as a selection idea; do NOT overcommit on two data points. ACTION TAKEN: add #357->yellow anchor variants to targeted sweep (yellow is independently over-determined). #423->recycle where it appears. ## D6 — Materials update (2026-08-27): anomaly ledger + new registers Built theories/anomaly_ledger.json (per-GAN, 797 records / 405 GANs, additive; 108 GANs still clean). Added to SOLUTION_SPACE: near-miss-homage glitch class, W22 image- association word register, W23 homage-target word register. STANDING RULE: every new anomaly appends to the ledger. The homage / image-association idea is parked for future exploration (the collector: "interesting... explore further in the future"). ## Claude visual validation — static pass complete (2026-08-27) All 512 pieces reviewed by Claude's eye on 13 mid-frame sheets (theories/ claude_visual_validation.json). Temporal types (splice/flash/speedup/reverse/brief-eyes) still need frame-strip pass. Headlines: tag families largely CONFIRM (backgrounds, costumes, invaders, giant-eyes 449-456, tributes, corrupted, black-bg); NEW misses found: #174 invader-form in robo drop, #253 robot-in-solstice, #412 pig, #276 quadruped, #474 FIGURELESS invader wallpaper, letter-eyes #470-472, three-eye + eye-grid censuses, second-head/echo family ~12 candidates beyond tags; dotted-bg tag mis-defined (family texture). All pending the collector's ratification — his eye remains authority. ## Frame-complete re-validation (2026-08-27, after the collector's method correction) the collector falsified the #474 'figureless' claim -> new STANDING RULE: all-frames review required before asserting any untagged find. Re-validated all 24 contested finds: RETRACTED 5 (474 figureless, 9 echo, 412 pig, 409/495 absence), REASSIGNED 3 (245-> letter-eyes[solstice!], 276->two-creature morph, 429->eye-multiplication), SURVIVED 15 (echo family 47/75/103/109/122/124/125; 253 robot-in-solstice; 174 invader-form; letter- eyes 470-472; 264 turkey; 268 alien-morph; 50 hammerhead-phase). NEW from corrections: #121 is a persistent TWIN at Dim 2 (slot-2 relevant!); candidate type 'two-creature morph' (276,268,495,50). Full record: theories/claude_visual_validation.json. ## D7 — Twin census RATIFIED + ENTANGLED class discovered (2026-08-27) the collector graded all 17 twin candidates on the all-frames page: TWIN (13/13 named confirmed): 73 (+partial third head = also Morai echo), 298 (inner eyes shared/merged), 59 (second head never fully forms but Twin positioning), 121, 124 (Ghost tag explains visibility; the two MERGE from frame 36 to 1 across the loop), 47 ("many frames of clearly defined twins"), 373 Twinvader, 379 Spin Twins (converge = interact; PLUS collector SEES LETTERS "AI" CLEARLY frames 24-30 — painted-letters channel!), 132, 139, 157 (robo tiling dialect), 323 (borderline, ghost disappearing accepted), 395. NEITHER -> NEW CLASS "ENTANGLED": 281, 226 (bodies consistently intermingled), 225 (maybe), 100 ("entangled before Pindar had language for it"; Primordial sloppier look). NAME VERIFICATION: #281 = "Entangled glitchGAN", #226 = "Entangled bitGAN" — the visual definition matches Pindar's naming class. 15 Entangled-named pieces exist (7 solstice, 2 ghost, 6 review): 217,222,226,244,260,263,281,411,436,450,465,473,481,489,499. QUANTUM THREAD (the collector): entangled is quantum vocabulary — qubitGAN #130 -> Entangled pieces -> #512 Quantum Glitch -> next collection (quantum computers). Twin=classical pair; Entangled=quantum pair (intermingled/shared state); Spin Twins=interacting (spin). ## D8 — Name-census discoveries (2026-08-27, meticulous full-collection pass) Censused EVERY word in all 512 chain names (name_word_census.json). Findings: 1. PALINDROME/nagaGAN class is EXACT and COMPLETE: 8 pieces, palindrome names, occupying ALL EIGHT 6-bit binary palindrome positions, with mirror-loop animation. Triple-coded. 2. TWIN name-family = exactly the 13 the collector ratified (Twin n=8 + Twinbot 2 + Bit-Twins + Twinvader + Spin Twins). Name channel and eye channel AGREE completely. 3. ENTANGLED n=15 — new class, quantum vocabulary, essences include key/solution/book. 4. Hapax names (used exactly once) = the Named pieces; a ready selection pool. 5. Decimal-palindrome piece numbers cross-referenced with anomaly-richness (#191 err0r carries 13 distinct anomaly types — the single most over-determined piece found). All added to ledger (2069+ records) and SOLUTION_SPACE Addendum 2 with new S1/S2/S3 entries and five new SELECTION vectors (V1-V5). ## D9 — GLYPH CHANNEL RULING (2026-08-28): CLOSED except one piece the collector graded all 17 glyph candidates. Result: 15 TEXTURE, 1 PAYLOAD, 1 unresolved. - TEXTURE (not payload): 207, 281, 196, 457, 459-463, 467, 470-473, 481. His reasons: "binary roboGANs have binary codes for faces — this looks normal"; "that's just the binary background"; "just the eyes"; "just the background". - PAYLOAD (confirmed): #379 "AI" frames 24-30 — the ONLY confirmed glyph payload. - #196 (my strongest new candidate): "nothing, ignore". RETRACTED. - #165 "unsure | bad frame choice?" -> my frame SAMPLING was at fault, not the finding; his known 8941 exists. Same likely for #207's "#4" (his own prior find, graded texture on my frames). ACTION: re-render both ALL-frames full-frame for him to locate. CONSEQUENCE: the letters/numbers-in-art channel is CLOSED as a selection vector. It is NOT a source of eleven pieces. Confirmed carriers: #379 + (pending) #165, #207. The criterion from Addendum 3 (does the glyph survive the creature?) is validated by his ruling — everything that survives the subject he called texture. METHOD NOTE: my 3-frame sampling repeatedly missed transient marks. Any future visual claim must use ALL frames (already the standing rule after the #474 correction). ## D10 — CORRECTION to my reasoning + the #165 signature (2026-08-28) collector'S CORRECTION (I had it wrong): I wrote "if hand-painted marks are a class, three is not the count that matters — eleven is." FALSE. His doctrine: "If we can find different types in different drops we may be able to MIX AND MATCH these things." A 3-member hand-painted set is a valid CONTRIBUTOR to the eleven, stacked with twins/inverted/ palindromes/etc. Never again require a single class to reach 11 on its own. #165 SIGNATURE (his description, Claude verified): the 8941 is NOT a static painted number. It is the SAME REGION ABOVE THE HEAD transforming through digit states across frames 12-21. Claude sees 4-5 glyph cells (vertical bar | box-with-slot | thin bar | solid bar) morphing over those ~10 frames, ABOVE the head — i.e. OUTSIDE the head crop my sweeps used. This is a TEMPORAL glyph, not a spatial one. SEARCH CONSEQUENCE: the correct target signature = a fixed region (often above/beside the figure) that persists across consecutive frames while morphing through glyph-like states. Single-frame and sampled-frame searches are structurally blind to it. ## D11 — Skull drop painted-mark hunt, first half (2026-08-28) Skull #1-#64 re-swept at EVERY FRAME, FULL FRAME (fixing the head-crop blindness), using #165's frames 12-21 as calibration and the collector's texture rulings as hard exclusions. RESULT: ZERO high, ZERO medium. One LOW (#12, frames 15-20, above head) flagged by the hunter itself as probably the creature's antenna/headgear degrading. 13 candidates examined at 3.5-9x and REJECTED with stated reasons (recorded in ledger so they are never re-hunted): 1,3,16,24,27,34,35,40,44,45,47,59,61. Notable: #45 morphs convincingly above the head but resolves to the square eye band; #24's dashes hold position while the creature morphs = background; #47's '|||' spans full width = wallpaper. => Skull #1-64 is CLEAN of the #165 signature. This is now a TESTED negative, not an artifact of cropping. Second half (#65-128) still running. ## D12 — Skull drop painted-mark hunt COMPLETE (2026-08-28) #65-#128 re-swept every-frame/full-frame. ZERO high. ONE medium: **#119** (frames 1-6, above head, detached, dark blue/black '7'-glyph + stroke + 2-3 dot cells, re-forms each frame while holding position, vanishes as creature fades). Claude verified present, localized, morphing = structurally the #165 signature. Hunter held at MEDIUM because it sits where a horn/hair tuft would. *** CONVERGENCE: #119 is one of only TWO pieces Pindar ever publicly flagged as puzzle- relevant — "(119/128) is only Odd number of Type E (typo or important?)... 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." The single skull piece bearing the painted-mark signature is a promise piece. Traits: ess `attract`, Dim 3, Flaming, NOT burnt, mint 193. NEW RULED-TEXTURE FAMILY: "chest emblem" — glyph-like cell clusters on the torso, recurs on 66,67,70,72,87,90,91,106,108,110,114,126 => family feature, not hand marks. 18 further regions examined+rejected (logged) incl. #117 (closest call: localized blue 'A'+block but resolves to detached ghost-face fragment with dashes running to frame edge). => SKULL DROP CLOSED except #119. Painted-mark carriers now: #379, #165, #207, +#119(med). ## D13 — #119 REJECTED (2026-08-28) the collector: "That is just an eye remnant." The #119 above-head glyph is an eye remnant, not a painted mark. SKULL DROP (#1-#128) IS NOW FULLY CLEAN of painted marks — a tested negative at every-frame/full-frame resolution. SCOREBOARD (honest): every confirmed painted mark in the collection was found by collector'S EYE (#165 8941, #207 #4, #379 AI). Agent sweeps: 0 confirmed finds, 3 rejected candidates (#196 'nothing, ignore', #119 eye remnant, plus 15/17 glyph candidates ruled texture). Method is now sound (every-frame, full-frame, texture-vs-payload criterion, logged rejections) but yield from automated hunting is so far zero. STILL UNSCANNED at correct resolution: #129-#512 (384 pieces) — robo, spooky, solstice, ghost, review. These were only covered by the FLAWED head-crop/3-frame sweep, which is provably blind to the #165 signature. ## D14 — Painted-mark persistence threshold (2026-08-28) the collector (approving the Phase V #129-512 hunt): "the painted mark needs to be visible in at least 2-3 frames, otherwise it's probably just transitory noise." RULING: a painted-mark candidate must persist across AT LEAST 2-3 consecutive frames (holding position, optionally morphing through glyph states like #165). Single-frame flickers are transitory noise — reject at hunter level, log the rejection. Consistency check: all three confirmed marks pass easily (#165 frames 12-21, #379 frames 24-30, #207 "#4" persistent). Applied as a hard gate in the Phase V hunt. Phase V design note: confirmed carriers #165 (slice robo), #207 (spooky), #379 (ghost) are left IN the hunt range as blind positive controls — a slice whose hunter misses its known mark has a suspect negative. ## D15 — #207 painted-mark LOCATION pinned by the collector (2026-08-28) On the all-frames location-review page, the collector graded #207 CONFIRMED and pinned it: frames = "52-7 is where it is most clear." (i.e. wraps the loop: f52,53,54,0..7 ~ 11 frames) region = "bottom left written in green on the orange body" reads = "4" CORRECTION to prior records: the mark is "4", NOT "#4". It is GREEN on the ORANGE body, BOTTOM-LEFT (not above the head like #165). Persists across the loop-wrap, passing the D14 2-3-frame gate easily. Extraction coordinate for later index-reads: bottom-left region, green glyph, frames 52-54+0-7. Ledger updated. #165 pin still pending the collector on the same page. ## D16 — #165 painted-mark LOCATION pinned by the collector (2026-08-28) On the all-frames location-review page, the collector graded #165 CONFIRMED and pinned it: frames = "11/12 - 19." region = "Top center above the head between two \"antennas\" of the roboGAN" note = "The thing that is most clear and sits longest is the number 81 at f12/14/15" KEY REVISION: the clearest, most stable painted content on #165 is "81" (frames 12,14,15), NOT the full "8941" carried in prior records. "8941" may be the morphing sequence across f11-19, but the LONGEST-HELD, CLEAREST reading is "81". Location: top-center above head, between the two antennas (roboGAN). #165 IS a roboGAN (robo drop 129-192). Extraction implication (Phase M, not now): if painted numbers are indices, the stable values are #165->"81" and #207->"4". Candidate reads: wordlist[81], wordlist[4], or piece pointers. Do NOT sweep yet — logged for the recipe pools. Ledger updated. Both painted-mark location pins (D15 #207, D16 #165) now complete. ## D17 — Phase V hunt: #379 false-negative + DURABILITY RESTRUCTURE (2026-08-28) CONTROL RESULT (blind positive controls #165/#207/#379 left in the hunt un-flagged): #165 RECOVERED (129-192 hunter, HIGH) | #207 RECOVERED (193-256 hunter, HIGH) | #379 MISSED (321-384 hunter logged only 2 rejections, never examined #379). => The 321-384 ghost slice "clean" verdict is VOID and must be fully re-hunted. The blind-control design worked exactly as intended: it caught a lazy sweep. COMPLETED & DURABLE ON DISK (survived a mid-run credit exhaustion): 129-192, 193-256, 257-320, 449-512. PARTIAL: 385-448 died at #432 (need 433-448). VOID: 321-384. NEW CANDIDATES surfaced (for the collector's adjudication, NOT yet confirmed): #184 MEDIUM (robo, crown ornament morphing through "4"), #179 LOW (robo, likely mouth-grammar), #414 MEDIUM (ghost, reads "13"/"I3"), #424 MEDIUM (ghost, reads "F"/"E"), #256 observation (blue squares top-right, off-silhouette, hunter graded out). DURABILITY SPINE (the collector's ask after credits ran out): phaseV_status.py recomputes phaseV_manifest.json from the small per-chunk finds files + a control check. Fresh session runs it FIRST to see done/partial/void/remaining without reading transcripts. Rules now standing for all hunts: (1) chunks <=16 pieces; (2) hunters write their painted_finds__.json INCREMENTALLY every ~4 sheets; (3) shared HUNTER_BRIEF.md holds the method so re-dispatch prompts stay tiny. Remaining chunks: 321-336, 337-352, 353-368, 369-384, 433-448 (80 pieces). ## D18 — #379 "AI" DISPUTED by thorough re-hunt (2026-08-28) — needs the collector's re-ruling The 369-384 re-hunt examined #379 carefully (per instruction, frames 20-32) and logged it in REJECTIONS, not finds. Its reasoned call: the text-like "AI"/"F" fragments appear at the BOTTOM EDGE of the frame where the creature silhouette crops the tiled GHOST-family wallpaper glyph (two square eye-holes + forked base), i.e. a wallpaper CROP ARTIFACT, not a painted mark. It found no independent letter/digit inside the twin-convergence itself. CONFLICT with D7/D9: the collector ratified #379 "AI" (frames 24-30) as the ONLY confirmed glyph payload. The hunter cites overlapping frames (24,26,28,31,32) but calls it texture. STATUS: this is a genuine texture-vs-payload disagreement, NOT a lazy miss (the hunter examined it and reasoned). New control state added to phaseV_status.py: DISPUTED (control in rejections w/ reasoning) vs MISSED (control in neither = lazy = VOID). #379 = DISPUTED. ACTION: put #379 on the adjudication page as a PRIORITY item with the competing readings + all frames, and ask the collector to re-apply the texture-vs-payload criterion (does the "AI" live and die with the creature, or persist as background?). Do NOT overturn D7/D9 — only the collector's eye can. New candidate from this chunk: #380 MEDIUM ("IE"/"1E", green on blue, chest). ## D19 — PHASE V ADJUDICATION COMPLETE (2026-08-28): #179 NEW CONFIRMED CARRIER the collector graded all 13 cards on the adjudication page. Verbatim export: #379 confirmed | #256 reject | #354 reject | #184 reject | #324 reject | #327 reject #362 reject | #380 reject | #414 unsure | #424 reject | #179 confirmed "maybe a 10 . 00" #364 reject | #441 reject RULINGS: 1. **#379 "AI" RE-CONFIRMED as payload.** The hunter's wallpaper-crop-artifact argument is OVERRULED by the collector's eye. D7/D9 stand. The dispute is closed. 2. **#179 is a NEW CONFIRMED painted-mark carrier** — the first agent-found, the collector- confirmed painted mark in project history. Region: the mouth-band position (hunter: hollow digit boxes, frames 16-31, robo drop). the collector's reading: "maybe a 10 . 00". Note the irony: the hunter itself graded this LOW and leaned "mouth-grammar texture"; the collector's eye saw through the caution. #179 is also the piece from the REJECTED pair {150,179} (D2: "179 is definitely a ghost — label fits") and the R2 label-swap idea. 3. **#414 held UNSURE** (ghost, reads "13"/"I3" f36-46). Not confirmed, not killed. Park. 4. All other candidates REJECTED: 256, 354 (the HIGH), 184, 324, 327, 362, 380, 424, 364, 441 — logged as adjudicated-rejected in the ledger; never re-hunt. PAINTED-MARK CLASS ROSTER (post-Phase-V, the class is now CLOSED census — all 512 scanned at every-frame/full-frame, skull clean D13, 129-512 adjudicated here): #165 robo -> "81" (D16, top-center between antennas, f12/14/15 clearest) #179 robo -> "10 . 00" (maybe) (mouth-band boxes, f16-31) #207 spooky -> "4" (D15, green on orange body, bottom-left, f52-7) #379 ghost -> "AI" (D7, f24-30, re-confirmed here) (+ #414 UNSURE only) SCOREBOARD UPDATE: agent hunting is no longer 0-for-N — Phase V recovered both blind controls AND produced one new confirmed carrier (#179). Method (every-frame/full-frame/ persistence-gate/texture-criterion) is validated end-to-end. EXTRACTION VALUES for Phase M (park, do NOT sweep): 81, 10.00 (or 1000?), 4, AI. ## D20 — Ledger reconciliation + STALE-STATUS correction (2026-08-28) Trigger: the collector caught #206 showing no glitch rating (its D1 theme-breaker ruling was never ledgered). A full DECISIONS-vs-ledger reconciliation ran: 164 piece-level claims checked, 141 already covered, 26 backfills reviewed and applied (all cite their D-ruling or validation-round source). IMPORTANT CORRECTION surfaced by the audit: the "Frame-complete re-validation" entry above lists #253 (robot-in-solstice) and #429 among SURVIVED finds — but the collector's later ratification round 2 REJECTED both ("not a robot"; "movement artifact"). Also rejected in rounds 1/2: #75 ring, #109 echo, #125 (ghost behavior); reassigned: #122 (nebulous), #495 (ghost-disappearance); corrected ID: #264 is a DOG, not a turkey. The ledger now carries the FINAL rulings; the earlier 'survived' language in this file is superseded. Also now ledgered: D2's {165,207} pair confirmation (with the 'suss' caveat), the {150,179} pair REJECTION, the TRUE/FALSE drop-relation interpretation on {363,473}, #204's forge-pair explanation, the D4 'matches' controls (#67,#179,#300), #117's closest-call rejection. PUZZLE-BOOK DATA POLICY (the collector's ruling, this session): the Collection tab is FACTS ONLY — dictionary confidence eye/confirmed/eye+byte gates the type filter and signal badges; heuristic/chat/agent material renders collapsed as 'Provisional — not facts'. Dictionary relabeled by provenance (42 types) + 41 new definitions; all 186 types defined. ## D21 — Chat-log mining + taxonomy + STANDING AUTO-UPDATE RULE (2026-08-28) FULL TRANSCRIPT MINE: all 28 sessions in this workspace searched; 156 puzzle-specific the collector messages reviewed one by one. 49 records backfilled with self-contained descs: - GLITCH-NAME LIARS (17, verbatim 2026-08-19 list): 9,15,25,35,45,46,51,56,58,60,72,75, 84,86,90,102,115 — named Glitch/Shimmer, no visual glitch. Mirror set: Shimmers that DO glitch {7,43,67,71} (= the byte-splice core; eye and bytes agree). Plus #166 Rainbot: glitches without the name, in a unique TWO-RHYTHM pattern (the collector 2026-08-20). - MONTAGE GRADES (2026-08-25): 19 mismatches + 2 '(?)' + 7 unsure + the #19 finding (its montage tile IS a glitch frame — f20==f40). Channel-as-selection died; grades preserved. - TRUE/FALSE earliest hypothesis (2026-08-07 verbatim) ledgered on #363/#473. TAXONOMY RULING (the collector): records split into GLITCH (confirmed, watchable, sortable) / OBSERVATION (descriptive cue for ideation — 'binary background' is not a glitch) / BOOKKEEPING. Dictionary carries a 'kind' field; the Collection tab now has TWO filters (confirmed glitches | traits+visual cues) and cards render three sections. SITE RULINGS (the collector): Selections tab = SELECTION HYPOTHESES (theories, incl. the new Montage-glitch-frame, PodGAN-invasion, Glitch-name-liars, Still-eyes, White-zero-eyes); Extraction/Ordering tabs = plain METHODOLOGY explainers (how each might work); the recipe BUILDER tab is KILLED (recipe.html retained on disk, off the site). DROP BOUNDARIES CANON (from the collector's own doubling note, fixes my wrong 320/444 bands): skull 1-128, robo 129-192, spooky 193-208, solstice 209-300 (92), ghost 301-444 (144), review 445-512 (68). Deck _drop_label is authoritative; never hand-code ranges. STANDING RULE (the collector): every new finding updates ALL materials AND the site, immediately. Mechanism: append to ledger/dictionary -> run ./update_all.sh (rebuilds workbench, pieces/sheet/methods/content/recipe data, stages the serve dir). No finding sits only in chat ever again. ## D22 — Tag-membership FACT CHECK + data corrections (2026-08-28) the collector asked whether the glitch categories/tags themselves are fact-checked. Built verify_tags.py (now step 5 of update_all.sh): re-derives every mechanically checkable tag family from raw data and compares against the ledger. All 9 families VERIFIED after these corrections: 1. **BURNT SET CORRECTED: 30 pieces, not 29.** burnt_gans_set.json was stale — the chain burn events (burnt_final.json, 32 events) include #66. True set: 2,9,11,20,32,33,49, 66,69,74,86,87,95,111,117,123,125,135,138,144,149,153,168,194,198,204,205,206,207,208. NOTE: promise piece #32 IS burnt. #66 (the chain-vs-Arweave SubType oddity) IS burnt. Phase N's earlier "burnt n==29 VERIFIED" checked against the stale file — superseded. 2. **'flag:No on-chain traits' RECLASSIFIED as bookkeeping** — it is a Blockscout CACHE artifact (its own desc says so); full chain records exist for all pieces. It was wrongly presenting as a confirmed glitch on 120 cards. 3. **Radix corruption finally tagged** on #465/#466 (was in DECISIONS/dictionary only). 4. **Frame counts now byte-authoritative**: frame_counts.json (all 512 measured from GIF bytes) feeds sheet.json; the deck's FrameCount_measured field had gaps. 5. **'Name contains BIP39 word' rule clarified**: it marks BIP words in PROPER-NAME (formula-breaking) titles only — 16 pieces; each claim spot-verified against chain name + wordlist. Formula vocabulary (ghost, smoke...) intentionally excluded. 6. Four weak/placeholder dictionary definitions rewritten (name_anomaly was literally "definition pending"). All 192 types now defined with kind + confidence. Eye-ruled tags cannot be machine-checked; verified instead that every collector_* record carries provenance (100% do). Readability rewrite of legacy record descs in progress (three editorial reviewers, proposals pending my review). ## D23 — RIGOROUS REVIEW: master audit + workbook doctrine + readability apply (2026-08-28) the collector: "do a rigorous review to make sure nothing we've talked about has been left out or catalogued wrong." Executed: 1. MASTER AUDIT (master_audit.py): 63 claims — every ratified roster, his verbatim lists (invader-shirt 28 + heads {465,466} + backgrounds {473-476}, pod-symbol 16, liars 17, shimmer-glitchers 4) name-resolved against exact chain names, pairs, byte facts, and D-ruling specifics. 4 gaps found and fixed (#205 D1-variant, pair {16,100} records, #511 D4-unsure, #32 promise+burnt). NOW 63/63 OK; 11/11 tag families verified. 2. DESKTOP WORKBOOK INCORPORATED (bitGAN_MASTER_annotated.xlsx) — MAJOR catalog gaps: - GLITCH GROUND TRUTH: the collector's corrected definition (one frame re-inserting with gaps shrinking 5,3,2,1,1 = backwards Fibonacci); detector matches his eye on all 21 hand-checks; 46 pieces glitch; **EXACTLY ELEVEN carry the complete run: 148,152,158,160,166,170,172,176,182,190,287** -> new selection SEL-FIB-ELEVEN, per-piece records, dictionary types, fact-check roster. Extracted to glitch_ground_truth.json. - PRIME RULE: skull glitch iff prime-and-not-Fibonacci (25/26); Pindar encoded the classes on-chain in unpublished HiddenType ('b' = glitchers). Exceptions recorded: #47 (should, doesn't — the Twin), #5 (shouldn't, does — the whale). - Old held-frame "glitchers" {231,256,274,474,479} are NOT recurrence-glitchers (kept as the separate byte-negative/timing classes they are). 3. READABILITY APPLY: 3 editorial passes over all ~2,400 records; 592 rewrites + 15 salvages applied (originals preserved in orig_desc); 262 misattributed chat fragments marked hidden (never deleted; site skips them). Root causes fixed: chat_mined records were stored 200-char-truncated AND rebuild_workbench.py truncated all descs to 200 for display — display truncation removed. 4. OPEN ITEMS PAGE for the collector: verify_open.html (nagas 147/159/162 still-eyes; #225 entangled-maybe; #370 Eye Swarm fuzzy; two un-numbered color-flash pieces; two un-numbered montage observations; glitch-vs-observation kind vetoes). REMAINING: workbook's other sheets (Candidate Elevens, What We Have Ruled Out, Burn & Forge partners, XMP Provenance...) being diffed vs catalog by agent — proposals to review. ## D24 — the collector's verification verdicts + FAMILY LADDER + workbook incorporation (2026-08-28) VERDICTS (verify_open export): #159 Dog God HAS still-eyes; #147 Taco Cat and #162 Star Rats do NOT. => STILL-EYES FINAL: {129,130,141,159,161,174,180,192} — eight pieces, six of eight nagas + qubitGAN + #161. CONCLUSION: NOT naga family grammar — a deliberately chosen class of eight. (#162, a naga WITHOUT it, has essence "twin".) Also: #225 parked permanently-unsure on entangled; #370 Eye Swarm CONFIRMED in the invader-shirt set; the two un-numbered claims (color-flash pieces, montage robo notes) DROPPED per the collector. TAXONOMY RULING (the collector): treat unique things as unique, small clusters as valid, AND ladder up to inclusive parents (miscase + transposition are each specific; both are Name Anomalies). Implemented: every dictionary type carries a `family` (12 families); the Collection filter groups by family with whole-family options. #194's vague name_anomaly record replaced with its exact defects (leet '8bit' cross-drop + selective double-space + formula skip). RECORD-QUALITY BAR (the collector): a record on a specific piece must state EXACTLY what is wrong with THAT piece — generic definitions belong to the type, not the record. WORKBOOK DIFF (agent, reviewed): ~120 claims; key incorporations — - #130 qubitGAN contract DESCRIPTION says "1024 animations" (all other robos: 2048). Digit-string channel no word sweep could see; 1024 = half the 2048 wordlist. Recorded. - Burn & Forge parent-pair table: 30 per-piece forge_pair records (each burnt piece's named partner + the SuperRare output from the artist's own Burned-bitGANs trait), including the pair burnt but NEVER forged — an outstanding on-chain debt. - HONESTY CORRECTION: the backwards-Fibonacci eleven's status downgraded from "ground truth" to "exact-11 (detector caveat applies)" — the Theories log warns the two glitch detectors disagree badly and NO eleven, including this one, is confident until resolved. The caveat is now stated wherever the class appears. - Claimed #192-HiddenType error CHECKED AND REJECTED: chain shows #192 HiddenType='F' (so does #193); the ledger was right, the sheet's self-contradiction is the artifact. REMAINING from the diff (to review next): Descriptions-sheet rows (14 more), Candidate- Elevens constructions (18), What-We-Have-Ruled-Out entries (5), rare-frame-eleven contradiction, File-Structure gaps (#188 canvas, #228 five-field outlier). COLOR SWEEP: deterministic scan of all 512 done; flash detector being calibrated against the known members {462,463,464,497} before candidates go to the collector. ## D25 — the collector's definitions-review calls applied + NORMALIZED TAXONOMY (2026-08-28) the collector graded ~95 legacy types and asked for ONE review surface. All calls applied: DATA CORRECTIONS from his eye: - **#5 does NOT glitch** — detector false positive; recurrence set 46 -> 45; the Prime Rule now holds 25-of-26 with #47 the ONLY exception (#5 follows the rule). - **#43 removed from larger-eyes.** - **Burnt clarified**: wallet holds 32 items = 30 Act-1 pieces + 2 from OTHER collections (Noun GAN Study 3/63; reneGANs 0110) — cross-collection burns fed forges. - Def fixes in his words: inverted = inserted frame is the NEGATIVE of normal frames; #204 forge-pair = explains the anomalous Zombie SubType (named for what it BECAME); chain-corrupt-side/corrupted-essence defs made concrete. NOTABLE STATUS CALLS: TWIN -> observation (his call — the ratified twin census stays, but twins are context, not glitches); missing-element -> GLITCH (absence-as-glitch); xmp:primes_missing and byte-identical pair -> GLITCH; resave sessions/burnt/description anomaly/invaders/figure-resemblance -> observation; shimmer/quantum-vocab -> bookkeeping. CUTS (15 types retired into a 'pending re-audit' bin, records preserved): glitch (hodgepodge; his note: #474 "doesn't glitch at all"), signature_violator, the chat/ collector color buckets (rainbow/glow/flaming), duplicate eye types, precursor, painted location/reconfirmed variants (folded into the one painted-mark concept). TAXONOMY v2: 96 concepts (from 199 types); statuses 47 confirmed-glitch / 4 candidate / 38 observation / 7 process; 44 concepts pre-approved from his calls; SIX flagged gold needing his wording (two-tier glitch def, montage-glitch-frame def, HiddenType-F def, name-misc 'Thre…' truncated note, missing-element-as-glitch definition, label-mismatch carve-up). Review surface: taxonomy_review.html ONLY — definitions_review retired. NOTE on his "I rely on you to confirm" items (xmp/file/chain claims): those families are machine-verified — Phase N re-derived them from bytes and verify_tags re-checks 11 families on every update. His reliance is backed by standing automated checks. ## D28 — Ran the open hypotheses; re-save sessions + glitch-geometry RULED OUT (2026-08-29) Ran the two computable open threads, plant-validated, pushed to the live site (Hypotheses #148/#149): - RE-SAVE SESSIONS as the eleven: 11 sessions (firsts 7,129,159,203,246,239,211,300,302,407,446; lasts 127,147,159,203,264,239,294,300,439,407,512). Piece-word reads (essence/piece#/mint#/name x 4 orderings) AND date-number reads (day-of-year/MMDD/month/days-from-first -> wordlist) all DEAD. - FIB-ELEVEN GLITCH GEOMETRY: position-sum / first-pos / last-pos / frame-count -> wordlist, DEAD. gap-sequence 5,3,2,1,1 is identical per piece (non-discriminating). Only eyes-as-binary (#150) remains. Open hypotheses now 7. #150 eyes-as-binary still blocked on eye-position extraction (needs Michael eye/detector). SITE: Selection tab REMOVED (Michael: tags+hypotheses cover it better). Nav now Puzzle/Collection/ Definitions/Extraction/Ordering/Hypotheses. ## D27 — Overnight sweep campaign + PRIME-POSITION DISCOVERY (2026-08-28 night) Authorized by the collector ("run down the queue... if negative, new hypotheses via recipes"), consideration set now stable (taxonomy ratified). All local/free, plant-validated. DISCOVERY: the Prime Rule extends to the ROBO drop on BINARY position (the (010011) suffix in each name), not decimal piece number. Every glitching robo sits at a prime-not-Fibonacci binary position (15 of 15); the only clean primes are 2,3,5 (the Fibonacci primes); zero non-prime glitchers. Same rule as skull drop, keyed to each drop's native numbering. The workbook missed it by testing robo PIECE numbers. Recorded on 15 robos as prime_position_glitch. Caveat: clean only for the two frame-splice drops; other drops glitch in other media and can't be frame-tested. SWEEPS (all NEGATIVE, plant-validated): fib11 (essence + all number registers + geometry × 7 orderings); chaincorrupt11 (exact-11); entangled11 (all 1365 subsets, essence×mint); quantummeta11, still-eyes+wilds, glitch-liars-11, painted+pairs mix; prime-15 robos (all C(15,11), essence + binpos). => essence and simple-number registers now DEAD on every ratified exact-eleven. Queues: q_night_1 (1581), q_night_2 (2730); logs *_w1.log. NEW HYPOTHESES (OVERNIGHT-BRIEF-2026-08-28.md): H-1 glitch's own number as the word (binpos swept dead; gap-sequence/repeat-count unbuilt); H-2 eyes-as-binary on white-zero pieces (needs eye-position extraction); H-3 the eleven = the eleven RE-SAVE SESSIONS not pieces (dates as words, Solstice-style, wholly unbuilt); H-4 prime-rule cross-drop selection; H-5 painted numbers as pointers (MAP component); H-6 Fibonacci-position skulls (HODL eleven) via number registers. Recommend building H-3 (sessions) and H-1 (gap-sequence) next. ## D26 — TAXONOMY RATIFIED (2026-08-28) + verification re-runs ordered The collector graded all 96 concepts. Applied: - 56 concepts RATIFIED. 19 CUT (incl. all process bins — process records stay in the ledger but get no taxonomy node; dupes folded into his approved originals: color-flash, scanline, morai-echo). Two-rhythm #166 = "a normal glitch" -> merged into frame splice. Frame-recurrence merged into frame splice (flagged for his confirm). Renamed per his instruction: Quantum-Glitch-Prelude (#501-504). Status flips: cross-drop hybrid, deliberate pairs, quantum thread -> observation. #205 REMOVED from theme-breaker (his catch: its D1-variant record is process info, not theme-breaking). - Taxonomy now: 72 concepts; 66 legacy aliases excluded from the taxonomy view. - NINE gold follow-ups remain (merge targets / status calls with no note: timing split, montage-glitch-frame home, hidden-element, radix home, CORRUPTED-type home, channel- banding, hiddentype, tributes, missing-files) + TWO piece decisions (#150 Firebot blue-stretch, #511 Inverted) now presented as proper cards with art. - MISSING-FILES ANSWERED + RE-VERIFIED: the Arweave store only ever covered 339/512; 173 pieces have no Arweave file (whole ranges). #134/#446 absences are from Pindar's own drive folders; display copies came from OpenSea snapshots. - VERIFICATION RE-RUNS (his order): the frame-recurrence detector REBUILT from scratch (the original script was never saved) and re-running over all 512 to independently re-derive the 45 glitchers + the backwards-Fibonacci eleven. Chain/metadata families already re-verified continuously by verify_tags (11 families) + master_audit (63 claims). NEXT once gold answers land: switch the Collection tab's filter from legacy types to the ratified concepts (the taxonomy becomes the site's single vocabulary). LORE FACTS mined for the record: (1) Pindar had no master plan — drops evolved over a year, cross-drop structure is retconned (the collector 2026-08-20). (2) Act 1 Review minted at the same time as the ghost drop; Pindar held all 68 and auctioned them later. (3) the collector authored the community 'portal glitch' puzzle (Fibonacci Caesar shift) — Pindar promoted it. (4) the collector believes GAN #0 was published AFTER the puzzle launched (verify on-chain before using #0 in any selection). (5) The prize GAN is a COLLECTOR tribute (Colborn, essence 'collect'), the counterpart of the artist GANs. ## D29 — Lore-catalog run-down + abnormal-HiddenType eleven (2026-08-31) Michael pasted the 46-item GAN-culture catalog (A1-A20, B1-B26, written 8/26, pre-D28) and asked to run everything down autonomously. Ran 10 parallel censuses + targeted plant-validated sweeps. Full triage: theories/LORE-CATALOG-TRIAGE-2026-08-31.md; consolidated findings + reports: lore_catalog_runs/FINDINGS-2026-08-31.md. NEW LEAD (selection, not a solve): the ABNORMAL-HiddenType eleven. HiddenType encodes each piece's number-class (skull a=Fibonacci, b=prime-not-Fib=glitchers, c=squares). The 11 pieces OUTSIDE the documented a-e/A-E scheme: F (9 = the 8 binary-palindrome robos {129,141,147,159,162,174,180,192} + spooky #193) and "32" (#32,#161). Clean, deterministic, on-chain selection of eleven, invisible to letter-only readers. SWEEP: DEAD on 16 registers (essence/pnum/mint/binpos/pos/name x mint/gan/binpos/rev orderings), plant-validated, registered. Mint-order essences advance alpha awkward cat dog news twin olive wolf zero glue were untested pre-today; now dead. STATUS: a ratified-QUALITY selection for the mix-and-match pool; NEEDS Michael's eye to endorse "out-of-scheme HiddenType" as a glitch class and for un-computable registers. CLEAN NEGATIVES this session (all catalog items, none land on 11): gallery archaeology (only pre-clue snapshot = known forge collection, plain order 1-11), journal images (clue post: no barcode/QR/marked word), walk-shapes, distinguished-index sequence, perfect-fakes/event-counting, odd-file bespoke reads (+ container census: disposal=none 100%, 0 comment/local-palette/trailing bytes on all 512), near-duplicates (new cross- drop pair #150<->#458 flagged), tile-scans, forge inheritance (never-forged pair = #49+#20; cross-contract anomalies=8 not 11), depth census (single-field template transfers to no field). CORRECTIONS to the catalog: "crispy" POSTDATES the clue (mint 2022-05-25) so not an Act-1 glitch; the "four skipped BIP39 words" claim is FALSE (only 2 skull-essence deviations, 1 unused word). Confirmed curios: 39-triple, 92-vs-93. PROTOCOL: registry audit confirms 100% full-128-ending coverage (no literal-ending gaps), BUT the BIP-39 PASSPHRASE dimension was essentially never run. Launched a passphrase sweep (universe/infinite as passphrase + universe as literal word-12) over 144 strong dead sets x 2 orderings — a genuinely new dimension. RESULT: 74,720 derivations, plant-validated, NO HIT. The passphrase/word-12 'universe' dimension is DEAD on these families. 'before the first, after the last' as a passphrase does not rescue the strong dead sets. Also novel for a future run: Union-eleven index-transplants -> mint-rank/montage/ title-alphabetical orderings. ## D30 — HiddenType number-class scheme FULLY DECODED + extension + bookend sweeps (2026-08-31) Michael asked to (1) extend HiddenType across the collection, (2) try anchor/bookend variants freeing 9- and 10-size glitch sets, (3) confirm the #150/#458 near-pair, and (4) test backwards-Fib ordering by first-insertion location. HIDDENTYPE DECODE (the field is Pindar's on-chain number-class encoding): a=Fibonacci b=prime(non-Fib) c=perfect-square d=odd-composite e=even Identical scheme across all THREE labeled drops, each on its native numbering: skull=decimal piece#, robo=binary position (with F=palindrome OVERRIDE class, the 8 nagas), spooky=octal position. Solstice/ghost/review carry NO HiddenType, so the scheme does not extend past spooky (no label = nothing to misplace). FOUR AUTHORED MISPLACEMENTS (label contradicts number-class) — the real anomalies: #32 literal "32" (skull) — Pindar-flagged promise piece #119 odd number in class e/even (should be d) — Pindar-flagged promise piece #161 literal "32" (robo) #187 even in class C/square (should be e) — essence 'mirror' => #32 and #119 are BOTH the hidden-types promise pieces; #161/#187 are their robo twins. This is a clean 4-member CONTRIBUTOR class (mix-and-match), NOT an eleven. It corrects D29's "F+32 eleven": F-in-robo is the legit palindrome class, not corruption. #150/#458 NEAR-PAIR: confirmed DISTINCT from byte-identical {14,75} and double-mint #390. #150=Firebot roboGAN (robo, ess fire), #458=review (ess soup). Best-FRAME match 150.f4 vs 458.f32 = cosine 0.085 (means were "close but not that close"; frames match much better) — a clean color-transformation, visually unmistakable. BUT the systematic "every Review GAN transforms a specific earlier GAN" hypothesis is NOT supported as a 68-wide map: no distance elbow, no earlier-drop clustering (matches spread ~ drop size), non-injective (55 targets for 68 pieces, GAN360 hit 3x). 458->150 is a real deliberate tribute in generic noise; the wholesale reprise-map is not there. review_transform_map.json. BACKWARDS-FIB LOCATION ORDERING (Michael's Q): first-insertion onset frame IS partly discriminating — 7 pieces share onset 37; #182=19,#287=22,#166=27,#190=29. Swept the new composite orderings (onset_then_mint, onset_then_mint_rev, drop_then_onset, onset_desc_mint) x {essence, piece#, binary-position} registers. plant-validated, NO HIT. Location as an ordering key is dead (glitch-onset was already in the D27 fib11 sweep). BOOKEND SWEEPS (Michael's anchor idea — anchors #0=universe, #1=abandon, #512=final): Tier 1 (64 curated configs): pairs-first9/last9 + {#0,#512}/{#1,#512}/reverse bookends; fib_robos10 + #512; palindrome8/still_eyes8 + {#0,#1,#512}. essence+piece#+mint registers x mint/gan/revmint orderings. plant-validated, NO HIT. Tier 2 (27,742 tasks, ~3.5M seeds): ALL C(n,9) and C(n,10) subsets of twin(13), entangled(14 usable), prime_robos(15), bookended by {#0/#1 .. #512} and {.. #512}, essence register, mint order. RESULT: 19/19 chunks, plant-validated, NO HIT. Tier 3 (1,270 tasks): ALL 512 one-member-per-pair combos (cross-drop, since pairs span skull/robo/solstice/ghost/review) + drop-native skull-11/solstice-11 trimmed to 9/10 + review-9 + spooky+ghost union, bookended. plant-validated, NO HIT. => the entire bookend campaign (T1 curated + T2 class-subsets + T3 pairs/drop combos) is DEAD on the essence register. The anchor structure does not rescue any ratified glitch set. Registered ~28k new keys. FRAME-BY-FRAME (Michael's Review-reprise ask): every frame of all 68 review pieces vs all 22,753 earlier frames. NO authored reprise set: best-frame distribution min 0.085 (that lone outlier = #458 soup <-> #150 fire, cosine 0.915), median 0.215; identity=1.000, adjacent-frame=0.972. #150/#458 confirmed coincidental family resemblance (Michael's read). PRIMER: published artifact 'The Hidden-Type Misfiles' explaining the scheme + 4 misfiles with embedded animations (lore_catalog_runs/hiddentype_primer.html). Pool + builders: lore_catalog_runs/glitch_pool.json, build_bookend*.py, q_bookend*. ## D31 — Global number-class scoring + full materials/site update (2026-08-31) Michael: (1) extend the HiddenType number-class principle to the WHOLE collection as one 0-512 sequence (not drop-native), record as properties; (2) profile the global-Fibonacci set. GLOBAL NUMBER-CLASS: scored every piece by its absolute GAN number (0-index) and mint index using the same classes (fib/prime/square/odd-composite/even). Recorded as deck properties GlobalNumClass_gan / GlobalNumClass_mint (513 rows) + pieces.json + dictionary prop:GlobalNumClass. Data: lore_catalog_runs/global_numclass.json. YIELD (honest): the drop-native prime-not-Fibonacci glitch rule does NOT extend globally (only 13 of 73 ratified glitch pieces are globally prime-not-fib); no ratified glitch class maps to a single global number-class. So the HiddenType scheme is DROP-NATIVE, not global. Global class counts (0..512): fib14 sq20 prime91 oddcomp146 even242. GLOBAL-FIBONACCI SET = {0,1,2,3,5,8,13,21,34,55,89,144,233,377} (14 pieces): - The first 11 (0-89) ARE exactly the HODL/Named/Fibonacci eleven (Pindar's own published eleven; the Named pieces). Carry the first BIP39 words by the skull essence rule. - The extra 3 (beyond HODL) each sit in a different later drop and each has an anomaly: #144 robo, ess flame, BURNT-in-forge (144=12^2, the unique square+Fibonacci; forge-linked); #233 solstice, ess basket, signature_violator / no-on-chain-traits; #377 ghost, ess drip, VISUAL SPACE INVADER + podgan_background (in the pod-invasion set). - NONE of the 14 are in the ratified VISUAL glitch classes (twin/entangled/palindrome/ painted/fib11) — the Fibonacci pieces are the Named/archetype pieces, a separate axis. - #21 (Fibonacci) carries the 'Essence out of sequence' flag (bean vs expected basket). - #5 (Fibonacci) is the prime_rule_exception (5 is prime AND Fibonacci -> protected). MATERIALS + SITE UPDATED (standing rule): ledger (HiddenType misfiles + selection), DECISIONS D29/D30/D31, dictionary (HiddenType defs enriched + GlobalNumClass prop), build_methods.py (new SEL-HIDDENTYPE-MISFILE, 22 selections), build_pb_content.py (prose), deck_512.json (GlobalNumClass cols), regenerated pieces.json/sheet.json/methods.json/ pb_content.json/taxonomy.json, registry (+~29k bookend keys). Site rebuilt: the workbench rebuild targeted a DEAD scratchpad serve dir (prior-session UUID) — built pieces.json directly instead (lore_catalog_runs/build_pieces_direct.py), fixed a live 'tc is not defined' bug in puzzlebook.html header, re-staged to this session's serve dir, repointed launch.json puzzle chdir. Verified live at localhost:8899: header renders, 22 selections incl. HiddenType misfiles, numclass in piece data. NOTE for next session: rebuild_workbench.py (SP=92176c78...) and update_all.sh (REV=d9ce74a2...) still hardcode dead scratchpad UUIDs; they need their paths repointed to the running session before update_all.sh will work end-to-end. ## D32 — "32 as building block" brainstorm swarm + Tier-A results (2026-08-31) Michael: brainstorm all ways the number 32 (the collection's building block) could decode the puzzle; then do both remaining Tier-A ideas. 7-lens swarm -> 49 ideas -> 24 deduped, tiered. Board artifact: lore_catalog_runs/brainstorm32_board.html; full doc BRAINSTORM_32.md. TIER-A RESULTS (all 5 top picks resolved): - ±32 hop: KILLED by control (20.0% landing vs shift-16's 30.6%; 32 not special). - doubling-eleven {0,1,2,4,8,16,32,64,128,256,512}: swept dead (essence+number x 4 orders). - pixel-diff 32x32 grid-cell -> 11-bit index (fib-eleven): swept dead (16 variants). The glitch is a re-inserted WHOLE frame, so the whole figure redraws -> no isolated signal cell (top1/top2 ratio 1.00-1.07). First pixel-layer read attempt; method doesn't fit this glitch. - bitGANg 1001 roster: DEAD END. It's a 22-for-2022 auction NFT = GIF of "32 unique bitGANs" (rainbow ghost faces), framed in lore as a sneak-peek at the NEXT drop's style -> NOT members of the 512; tile-to-deck shape match ambiguous. Intersection not well-defined. - {#32,#95} burn pair: VERIFIED real (same-tx burn 2021-10-14T14:49:06); candidate for D2's 10th-pair slot. Essence read as 10th-pair+anchor (first/last-speaks x 357/512 x mint/pairorder) swept dead. The PAIR stands as a structural finding; its word-read does not. NET: 4 of 5 Tier-A picks negative on word-reading; the {#32,#95} pair is the one durable new observation. Structural keeper: #129<->#161 (both still-eyes, differ by exactly the 32-bit). Remaining open (board): mint-as-pointer-nonskull, 16-blocks-of-32 histogram (extend), base-32 radix, 64-page split, 32-based reading orders (pair w/ non-essence register), 32-row bitstring, distance-from-32, base32-text. Consistent lesson: 32 is a real structural axis but every 32-based WORD-READING tried is dead — same essence-register death seen project-wide. ## D33 — 32 as selection + base-system analysis (2026-08-31) Michael: can 32 PICK the GANs? And how does 32 sit in base-32 / octal / binary? SELECTION (does not isolate 11): GAN mult-of-32 = 16-17 pieces (4 glitch); no residue mod 32 lands on 11 or concentrates glitches; 16-blocks-of-32 FAILS because glitches cluster in the robo drop (blocks 4-5 hold ~33 of 73), not one-per-block. Uniform 32-partition can't select eleven. BASE SYSTEMS: - 32 = 100000 bin = 2^5 (6th bit) = 0o40 octal = "10" base-32. - Base-32 exists (duotrigesimal / RFC-4648 Base32). Every GAN 0-511 = two base-32 digits (hi = block 0-15, lo = position 0-31) = the 16-blocks structure; glitches cluster in hi-digit 4,5 -> not a clean base-32 selector. - Drop sizes: skull 128=2^7, robo 64=2^6, spooky 16=2^4 ARE powers of two; 2^5=32 is the conspicuously MISSING drop size while being the universal 32x32 building block. bitGANg 1001 is a 32-set but not a drop. - KEY INSIGHT (explains why exactly TWO "32" corruptions exist): each of the first three drops has its own base and its own position range where a "32" could be planted - skull decimal 1-128 (32 valid -> #32), robo binary 0-63 (32 = 100000 valid -> #161), spooky octal 0-15 (32 = 0o40 OUT OF RANGE -> impossible). So 32 is plantable only in skull and robo, and both carry exactly one "32" -> the count of two is forced by the base structure, not arbitrary. - #161 = binary 100000 sits at the robo drop's exact top-bit midpoint (boundary between lower-32 pos 0-31 and upper-32 pos 32-63); it is simultaneously the "32" corruption and the base-2 midpoint marker. But robo glitches split 17 lower / 16 upper - near-even, no selection signal from the split. VERDICT: 32 is a genuine cross-base BOUNDARY / building-block (2^5, the missing drop, the robo bit-split, the base-32 cell) and the two "32" pieces are base-aware position markers - but 32 does not act as a selector that isolates eleven. Structural, not selective (same pattern as its word-reading death, D32). ## D33b — "32" extended to all 6 drops + 0-vs-1 indexing (2026-08-31) Native numbering per drop (from on-chain names): skull decimal /128 (1-indexed), robo binary 6-bit (0-indexed), spooky octal 0o (0-indexed), solstice ABSOLUTE #209-300, ghost decimal /144 (1-indexed), review ABSOLUTE 445/512. Two families: per-drop counters (skull/robo/spooky/ghost) vs absolute global numbers (solstice/review). WHERE "32" LANDS per drop's own system, and glitch status: - skull 1-idx 32/128 -> #32 (GLITCH: corruption + promise piece) [0-idx -> #33 clean] - robo 0-idx bin 100000 -> #161 (GLITCH: corruption + still-eyes) [1-idx 32nd -> #160, also glitch/still-eyes/fib] - spooky -> none (max 0o17 = 15) - ghost 1-idx 32/144 -> #332 'spider' (CLEAN, gc0) [0-idx -> #333 'trophy' clean] - solstice absolute -> NO native '32' (32nd solstice = #240 'before' clean; #241 'begin' clean) - review absolute -> NO native '32' (32nd review = #476 'unique' clean; #477 'december' gc1) FINDINGS: 1. Extending "32" to the other three drops adds NO new glitches. The "32 = a glitch" pattern is confined to skull + robo — exactly the two drops that have BOTH a HiddenType field to corrupt AND a position 32 that fits. ghost has a native 32 (#332) but it's clean (no HiddenType field there to corrupt); solstice/review use absolute numbering so a per-drop "32" doesn't even exist. 2. The two real corruptions use OPPOSITE indexing: skull #32 is 1-indexed (32/128), robo #161 is 0-indexed (binary value 32). There is no uniform global "32" - it is base- and index-relative to each drop. Under global absolute numbering, "32" = #32 only. 3. Mild curiosity (clean, likely coincidence): solstice's 32nd/33rd pieces #240/#241 carry 'before'/'begin' (beginning vocabulary). Noted, not load-bearing. VERDICT: the native-"32" pointer set = {skull#32, robo#161, ghost#332} - only 2 of 3 are glitches, doesn't build toward eleven. 32 remains structural, not a selector, across all six drops and both indexings. ## D34 — Untested-methods swarm + mirror-image search (2026-08-31) Michael: (1) do any Review GANs mirror earlier GANs (even one frame)? (2) swarm the website's UNTESTED selection/extraction/ordering principles against our uncovered glitches. MIRROR SEARCH (task 1): NEGATIVE. Every review frame's horizontal flip vs all 22,753 earlier frames, with flip-gain + self-symmetry controls. Strict gate (mirror<0.05 AND flip-gain>0.03 AND beats self-symmetry): 0 pieces. Max flip-gain anywhere +0.024 (flipping barely helps). Close matches = washed-out noise frames snapping to hub frame #77. No Review piece is a deliberate mirror of a specific earlier piece — generic pixel-art symmetry only. Report: lore_catalog_runs/mirror_report.md. METHODS SWARM (task 2): 4-family workflow, 28 recipes over 14 untested word-registers + calendar/gallery orderings x every glitch set. 4 Tier-A sweepable sets; on VERIFICATION: - A1 radix-decode on fib-eleven = ALREADY DEAD (re-derives the binary-position register; both index variants in registry). Caught before sweeping - the "most promising" was a repeat. - A2 tribute-11 (new selection: 14 person-tribute pieces minus 3 founder-anchors = {210,211,228,246,264,282,447,448,482,445,446}; words champion/escape/potato/mask/acid/ recycle/cloud/elder/test/prize/flower): SWEPT DEAD. Caveat: 5 of 11 words rest on external artist-bio guesses + #447's word = its essence, so a miss doesn't fully close the register - the tribute-word register needs Michael's eye on the actual words. - A3/A4 Silver-SubType-11: SWEPT DEAD (degenerate, 1-4 distinct words). - B1 TWIN-13->11 classifier (EC40): #157,#298 are the only 2 twins that invert Twin into Type instead of SubType -> a clean mechanical 13->11 cut. Swept the 11-set (essence + number registers): DEAD. Selector is clean but needs a live word register. Full recipe doc + tiers: lore_catalog_runs/METHODS_RECIPES.md. RESIDUES NEEDING BUILD OR MICHAEL'S EYE (not mechanically sweepable, genuinely uncomputed): - S2-W16 filename-descriptor words: 72 pieces carry an Arweave hand-written filename word (ancient/elder/heart/warrior/whale...) appearing nowhere in their metadata - a real uncomputed register; worth building a per-selection extraction. - S2-W13 tribute-word register: needs Michael's word confirmations (6/11 solid). - S2-W17/OC17 date/calendar decode on ENTANGLED (10/15 resolved); EC45 negative-space ('action','seven' the only unused words - 2 total); S2-W21 absence (7 words, short of 11). NET: no hit. The swarm's value is a verified map of the non-essence register space: most untested registers either collapse into already-dead ones, fall short of 11, or need semantic judgment. The cleanest genuinely-uncomputed mechanical register left is S2-W16 (filename descriptors). ## D35 — Filename-descriptor register + TWIN-inverter pointers + tribute eye-need (2026-08-31) Michael: "go and do these things" (the 3 residues from D34). 1. S2-W16 FILENAME-DESCRIPTOR register: extracted BIP39 filename words not in metadata -> 92 pieces carry one, but they are overwhelmingly SKULL, so only the Fibonacci/HODL selections are reachable (robo/palindrome/still-eyes files are just binary, 0 coverage). HODL descriptors are dominated by generic "skull"; swept the skull-descriptor set and the creature-hybrid set (creature-word where BIP39 else essence): DEAD, plant-validated. FINDING: the DISTINCTIVE filename words are the Named pieces' true creature identities - #5 whale, #13 spirit, #34 heart, #55 warrior, #1 ancient (all BIP39) - a real "true-class / creature-word" register, but only 5-6 of 11 HODL members have a BIP39 creature word (charred/scholar/blob/beanie/shark are not BIP39), so it cannot fill eleven alone. Parked as a genuine partial register (data: lore_catalog_runs/filename_descriptors.json). 2. EC40 TWIN-INVERTERS as pointers: #157 (SubType Binary/Type Twins, ess lonely, robo) and #298 (SubType Slender/Type Twins, ess brick, solstice) are the only 2 twins that invert. Pointer reads: #157 lonely->idx1053 (out of GAN range); #298 brick->idx221=#221 (entangled, a weak coincidence); GAN/mint pointers land nowhere marked; diff 141, sum 455. NO clean pointer. Dead as a pointer mechanism. 3. S2-W13 TRIBUTE register: the mechanical 6-solid-word version already swept dead (A2, D34). The 5 uncertain slots need Michael's eye (see report). Staged as an eye-question, not fabricated. NET: no hit. Filename register + creature-hybrid + twin-inverter pointers all dead. The one genuine partial that survives is the creature/true-class word register (5-6 BIP39 of 11) - which points back, again, at the true-class/Named-relation register as the real open frontier. ## D36 — NO-GUESS / MUST-BE-OBJECTIVE doctrine (Michael, 2026-08-31) Michael's ruling on the tribute-word register: "If we have to guess the words, that probably won't work unless you can see them in the image." Formalized as a standing principle that RETROACTIVELY invalidates a whole class of registers: THE WORD MUST BE OBJECTIVELY PRESENT — readable by anyone from the piece itself — via one of: (a) a metadata FIELD value (essence/subtype/type/name/date/hiddentype/filename), or (b) a mark actually RENDERED in the art (painted number/letter, an inscription), or (c) a visually UNAMBIGUOUS identity (you can SEE it is a whale -> 'whale'). It must NEVER require inference from outside knowledge or interpretation. This matches Pindar's "no judge" rule (the seed is objectively there) and his three solved puzzles (all mechanical: a field, the clue words, painted frames). CONSEQUENCE — these untested registers are DEAD by this principle (they require guessing): S2-W13 artist-tribute words (needs artist-bio knowledge: Abosch->potato etc.), S2-W23 homage-target word (needs interpreting what a piece "half-becomes"), EC46 GAN-vocabulary/lore-thematic words (pure interpretation), and the guess-dependent slots of any semantic register. WHAT SURVIVES as objective+visible: (1) the 4 confirmed painted marks (#165 '81', #207 '4', #179 '10.00', #379 'AI') - already a closed census; (2) the creature-identity true-class register where the creature is visually unambiguous AND a BIP39 word (whale/spirit/heart/ warrior/ancient) - but only 5-6 of 11 Fibonacci pieces qualify, so it can't fill eleven alone. The exhaustive all-frames painted-mark census (Michael-adjudicated) already answered the "is a word visibly rendered?" question for all 512: only 4 carriers, NONE are tribute pieces. => The tribute/homage/lore-inference registers are closed. The objective-word frontier is narrow: painted marks (4) + visually-unambiguous creature words (partial). Sweeps that require a guessed word are not worth running. ## D37 — Objective-word frontier: #300 / creature census / painted marks (2026-08-31) Michael: "do all of those" (the 3 objective-word directions). 1. #300 BYTEGAN DEEP-READ: NEGATIVE. Full byte/XMP/8BIM extraction of #300 "Lunar skullGAN": no Photoshop layers, no history, no 8BIM/PSD fragment, no seed keywords, no ordered BIP-39 run, no 11/12-item list. PREMISE CORRECTED: #300 is NOT the lone PSD file — ALL 512 carry a Photoshop CreatorTool, 16 share exactly "23.0 Windows", 511/512 have DerivedFrom; #299/#301 are structurally identical. #300's only uniqueness = an OriginalDocumentID save-path fingerprint, not a designed payload. byteGAN rhyme dead. 2. VISIBLE-CREATURE CENSUS (objective, D36-compliant): built for 511 pieces (26 creature words, 17 BIP-39, coverage 459/511). KEY FIND: exactly 11 RARE (<=2 occurrence) creatures are BIP-39 words -> an objective, visible, non-essence, exactly-eleven candidate: whale(5) spirit(13) heart(34) warrior(55) cat(147) dog(159|276) wolf(180) clown(204) turkey(264) alien(268) pig(412). Word set FIXED, only order + dog-piece vary. SWEPT 17 orderings (mint/gan/alpha/pos-in-drop/drop-mint/ess-alpha/wordidx/reverses, both dog variants), plant-validated: ALL DEAD, including mint order (the solved precedent). Strong negative — if this were the selection, mint should have hit. Data: lore_catalog_runs/creature_census.json. Slender-family is the other exact-11 creature set (11x'slender', degenerate). 15 one-off creatures catalogued. 3. PAINTED MARKS as pointers/indices: #165'81'->bip39[81]=any / GAN#81; #207'4'->above / #4; #179'10.00'->access / #10 (ambiguous 10/100/1000); #379'AI'=letters, NOT numeric. Only 3-4 objective items, one non-numeric -> a stacked COMPONENT of <=4, no complementary objective 7-register alive to pair with. Pointer-targets (#81,#4,#10) are all Silver/ Flaming skulls. Closed as too-small component. NET: no hit. The objective-word frontier (the space D36 leaves open) is now largely exhausted: #300 dead, painted marks too small, creature-11 (the best objective exact-eleven) dead across all principled orderings. RESIDUE FOR MICHAEL'S EYE: the creature-11 selection was built from filenames + prior validation, not his direct adjudication — if a creature ID is wrong (#268 alien? dog 159 vs 276? exactly these 11?) the set shifts. Worth his eye before fully closing, since it is the cleanest objective exact-eleven the project has found. ## D38 — Creature-11 CLOSED by Michael's eye (2026-08-31) Michael adjudicated the Creature Review artifact. Corrections: #264 (census 'turkey') is a DOG — but it is the Killer Acid tribute (cryptoArtGAN, name 'Killer acidGAN'), not a rare creature; #268 (census 'alien') is a BIRD — name is literally 'Bird monsterGAN' ess 'bird'; #412 (census 'pig') is a MONSTER — name 'Ultra monsterGAN'; #276 (a dog candidate) is NOT a creature (Nebulous bitGAN / Type Cloud) -> dog = #159 confirmed; #50 not a creature. CONSEQUENCE: the filename/visual census MISLABELED 3 solstice/ghost pieces. The clean "11 rare one-off BIP39 creatures" was a census ARTIFACT: #264/#268/#412 are actually a tribute and two monsters (monster is a common Type, not a rare creature). The genuine rare one-off BIP39 creatures number ~8: whale(5) spirit(13) heart(34) warrior(55) cat(147) dog(159) wolf(180) clown(204) -> 5 Fibonacci-Named + 3 nagas (Taco Cat/Dog God/Flow Wolf) + clown. Not eleven. Swept Michael's CORRECTED 11-piece reading anyway (whale spirit heart warrior cat dog wolf clown dog bird monster, 7 orderings incl mint): DEAD. Both the census version (17 orderings, D37) and the corrected version (7 orderings) are dead. => The visible-creature word register does NOT yield a valid eleven. The objective-word frontier that D36 left open is now EXHAUSTED: #300 dead, painted marks too small (<=4), creature register doesn't produce 11. STRATEGIC CONCLUSION: the entire "select 11 pieces + read each via register X" space is comprehensively closed (essence, all number registers, bookends, 32-everything, passphrase, pixel, mirror, all untested word-registers incl objective ones). This strongly implies the seed is NOT distributed across 11 individually-selected pieces. Next frontier = the phrase-lives-in-ONE-artifact shape (byteGAN precedent: 12 words in one file/composite): inventory + byteGAN-read every composite Pindar published (montage/all.jpg, the finaldrop & review grids, bitGANg-1001 rainbow-ghost image, the clue-post images), and re-open the gallery-arrangement channel with fresh archive angles. ## D38b — Creature register EXHAUSTIVELY closed (2026-08-31) Per Michael "try all iterations + essence words + orderings": swept all 128 combinations of creature-word-vs-essence-word for the 11 pieces {5,13,34,55,147,159,180,204,264,268,412} x 7 orderings (mint/gan/alpha/reverses/pos-in-drop/drop-mint) = 512 unique sets, 65,536 seeds, plant-validated: DEAD. (Note: cat/dog/wolf/bird have essence==creature; the naga trio + the Bird monsterGAN.) Combined with D37 (census-11, 17 orderings) and D38 (corrected- 11, 7 orderings): the visible-creature/essence register on this selection is COMPREHENSIVELY dead across ~530 combination-orderings. FINAL: the objective-word frontier (D36) is exhausted, and with it the entire "select 11 pieces + read each via ANY register (essence/number/creature/filename/painted) in ANY ordering" space. No hit anywhere. This is the session's decisive structural negative. NEXT (recommended): abandon the distributed-11 shape; pursue the byteGAN "one artifact holds all 12 words" shape — byteGAN-read every composite Pindar published (montage/all.jpg, the finaldrop & review grids, the bitGANg-1001 rainbow-ghost image, the clue-post face GIFs) and re-open the gallery-arrangement channel. ## D39 — PINDAR CONFIRMS metadata errors are intentional glitches (Discord 2022-01-01) Michael found a verbatim pre-clue Pindar message. @Krum: "All bitGANs have Type and SubType. The SubType only for this one is mis[sing]…" (re: a RAINBOW skull). VanArman (Never DMs), 2022-01-01 5:16 PM: "There are some glitch bitGANs with intentional errors - you will find others - they are glitched." IDENTIFIES: Krum's piece = #58 "Rainbow Glitch skullGAN" — chain has Type='Rainbow' and NO SubType (the SubType-value 'Rainbow' displaced into Type, SubType field emptied). Also #383 "Space-Clops" has the trait KEY misspelled 'Subtype' (lowercase t). UPGRADE: this confirms METADATA/trait errors as an intentional GLITCH channel ("they are glitched") with MORE to find ("you will find others"). It ratifies the CHAIN-CORRUPT-11 = {16,17,44,48,58,62,66,81,173,281,309} — exactly eleven pieces where the on-chain trait is the corrupted side — from candidate to PINDAR-CONFIRMED intentional-glitch class. #58 (Krum's) is a member. Status: SEL-CHAIN-CORRUPT-11 -> confirmed-by-artist. THE ERROR TAXONOMY (Michael's "divine why" — each error is a distinct type): #16 TRANSPOSE (Primoridal); #17 SUBSTITUTE title (Ghost->Glitch); #44 SUBSTITUTE title (Ancient->Primordial); #48 REVERSE (Rainbow->wobniaR); #58 DELETE/DISPLACE (SubType gone); #62 SWAP (Type<->SubType); #66 SUBSTITUTE (Rainbow->Flaming); #81 DELETE suffix (skullGAN->GAN); #173 SUBSTITUTE type (Ghost->Smoke); #281 SUBSTITUTE gang (glitchGAN->bitGAN); #309 DELETE suffix (ghostGAN->Ghost). ROSETTA STONE: the THREE Rainbow pieces show the same value 'Rainbow' glitched three ways — #48 reversed (wobniaR), #58 deleted, #66 substituted (Flaming) — a teaching device for the glitch grammar. EXTRACTION STATUS: essence register DEAD (D27). Corrupt-values BIP39 only 2/11; correct-values 3/11; reversed-essences 0/11 — so the word is NOT in the trait values. Per Solstice ("they hide their true BIP39 nature elsewhere"), the BIP39 word for these confirmed glitches lives in an unread channel. Most promising untested register: EC28 glitch-as-INSTRUCTION (the corruption names the operation: reverse/delete/swap/substitute). ## D40 — Glitch-as-instruction (EC28) on the Pindar-confirmed corrupt-11: SWEPT DEAD (2026-09-01) Michael: treat each error type as a VERB (transpose/substitute/reverse/delete/swap) and ask what OPERAND it acts on (essence, BIP# 0/1-based, GAN, mint, name words, trait values). Full menu enumerated (lore_catalog_runs/instruction_menu.json): ~90 candidate words across 11 slots; raw product 2.4e9 — the tell that a real cipher needs ONE consistent operand. Consistent-operand schemes: BIP-index families FAIL on #66/#281 (their 'substitute' has no BIP-39 true value: rainbow/glitchGAN); trait-value family fails on most. Only three fill all 11 slots: GAN/0b (12 combos), mint/1b (36), mint/0b (36). All 168 word-sets (mint + reverse orders) swept, plant-validated: DEAD. The three canonical phrases were word-salad (no authored texture), consistent with the negative. CONCLUSION: the corruption verbs are NOT digit-operations applied to the piece numbers. 'substitute' (5 of 11) has no clean numeric semantics and is what breaks the BIP families. The surviving reading: the verbs describe how each TRUE word was HIDDEN (reversed/deleted/ substituted) — recovery = UNDO the operation on an operand not yet identified — rather than applying the verb to a number we already hold. The corrupt-11 remains Pindar-CONFIRMED as the glitch set; its word-register is still open. ## D41 — Undo-the-corruption POINTER reading on corrupt-11: SWEPT DEAD (2026-09-01) Michael: undo each corruption to its TRUE value, treat that value as a pointer. Undone values: #16 Primordial, #17 Ghost, #44 Ancient, #48/#58/#66 Rainbow, #62 Nebulous, #81 skullGAN, #173 Ghost, #281 glitchGAN, #309 ghostGAN. FINDING: they point at SETS not pieces — Rainbow=82 pieces, Ghost=105, skullGAN=167, Primordial=29, Nebulous=19, ghostGAN=64; only glitchGAN=8 and Ancient=8 (filename) are small. Only Ghost/Ancient are BIP39 words. Collapse rules swept (9 schemes, 16 sets incl reverses, plant-validated): set-count->word (1b/0b), first-member essence, Nth-occurrence-member essence (GAN & mint), last-member essence, direct-BIP39-else-count hybrid, first-member-GAN->word (1b/0b). ALL DEAD. CUMULATIVE on the Pindar-confirmed corrupt-11: essence (D27), trait values (D39), instruction-on-numbers (D40), undo-pointers (D41) — all dead. The selection is the most artist-ratified eleven in the project; its word-register remains unidentified after four independent register families. Words are not in its metadata, its numbers, or the sets its true values name. ## D42 — Verb-by-LOGIC matrix + the missed GAN@1b logic swept (2026-09-01) Michael: don't mix operands (Pindar wouldn't); show every reading grouped by verb AND by operand-logic so picking one logic is visible across all eleven. Built lore_catalog_runs/glitch_verbs_by_logic.html (Part 1 by verb w/ full name, corrupted trait, correction, operand matrix; Part 2 by operand logic w/ the 11-piece phrase per logic). LOGIC FILL COUNTS (a logic is sweepable only if it fills all 11): BIP#1b→1b/0b 6/11, BIP#0b→1b/0b 6/11 (the five 'substitute' pieces have no numeric op on a BIP#), essence- letters 0/11, trait-value 6/11, GAN→word@1b 11/11, GAN→word@0b 11/11, mint→1b 11/11, mint→0b 11/11. CORRECTION to D40: the consistent-scheme code excluded any result equal to the piece's own essence, which silently dropped the GAN→word@1b family (for skulls, 'substitute index with GAN' reproduces the essence). D40 therefore swept only GAN@0b + mint@1b/0b. GAN@1b is a real, distinct, previously UNSWEPT full logic: always acoustic air apology {above|abstract} {add|amateur} amused {abandon|abstract} better cart {absurd|admit|age} — 48 combos (mint + reverse). Swept, plant-validated: DEAD. NET: all four operand logics that fill eleven (GAN@1b, GAN@0b, mint@1b, mint@0b) are now swept dead; the six that don't fill cannot form an eleven. Under the constraint 'one operand, applied consistently', glitch-as-instruction on the corrupt-eleven is fully exhausted. ## D43 — "Glitches TEACH the method" swarm: recipes swept (2026-09-01) Michael: apply the corrupt-eleven insight (a glitch is a lesson demonstrating a METHOD applied elsewhere) across every glitch class. 6-lens swarm -> 41 recipes -> TEACHING_RECIPES.md. META-PATTERN (all lenses converge): the glitches never hide the word in a piece's own essence/number (that space is closed); they teach INDIRECTION — a correct index/rank/value aimed at the WRONG list, column, or direction (podGALLERY's "right index, wrong list"). The most consistent TARGET is the BIP-39 wordlist itself approached sideways (reversed, re-based, transplanted), not any piece's own field. Secondary: structural glitches (splice gaps, held frames, palindrome mirroring) may encode the phrase's SLOT ORDER, not words. TIER-A SWEPT (all plant-validated, DEAD): A1 index-transplant (corrupt-11 mint-rank -> title-alphabetical rank -> essence: goat fox bullet cross air grab goddess blind eight divide bean) + my two independent recomputes; A2 reversed-wordlist read of fib11 binary-suffix N-values (work wonder wisdom winter win width whisper wheel welcome weapon swing — a list- MIRROR, distinct from the banned letter-reversal family) + recomputes on GAN/GAN-1/mint. 19 sets incl reverses. VERIFICATION NOTE: independent recompute reproduced A1 to within one sort-tie word and confirmed A2's operand = robo binary position read from the list's END. NEW FINDING from the swarm: the deck-wide typo/deliberate-pair census lands on EXACTLY 11 pairs = 9 ratified + {9,73} + {170,175}; #175 sits in two pairs. One-per-pair elevens (essence, mint order) being swept now (D43b). ## D43b — The 11-pair eleven swept DEAD (2026-09-01) Deck-wide typo/deliberate-pair census = exactly 11 pairs: the 9 ratified {14,75}{479,480} {209,300}{489,491}{130,133}{165,207}{363,473}{16,100}{175,205} + two new {9,73} (shared filename typo) and {170,175} (shared name anomaly). #175 sits in two pairs, so combos picking #175 twice were excluded (not eleven distinct pieces). All 1,536 distinct one-member-per-pair elevens, essence register, mint order: 196,608 seeds, plant-validated, NO HIT. Dead. The 11-pairs count remains a real structural fact (recorded); its essence read does not open. ## D44 — DELIBERATE-PAIR glitch class created + site pairs view (2026-09-06) Michael: make the pairs a glitch class and show each pair TOGETHER on the site with WHY they are paired, everywhere pairs are the glitch. CLASS (pairs.json, canonical): deliberate_pair = two pieces carrying the SAME one-off anomaly that appears on exactly those two in 512. ELEVEN pairs: 9 ratified (D2) + 2 new (D43b census, data-verified 2026-09-06): P10 {9,73} both Arweave filenames spell 'gltich' (Glitch transposed; #73's Type field carries it too) — #128's 'raindow' is a DIFFERENT typo, excluded; P11 {170,175} both SubTypes spell 'Raibow' (Rainbow minus n). #175 sits in two pairs (P09 'Gan' suffix + P11 'Raibow') -> 21 distinct pieces, 22 pair-memberships. Per-pair REASONS recorded (byte-identical / held-frame+palette / day-night / scanline / same flash / painted numbers (suss) / TRUE-FALSE / Primoridal / Gan / gltich / Raibow). MATERIALS: ledger deliberate_pair records (partner+reason) on all 21 pieces; dictionary type deliberate_pair (kind glitch, eye+byte, family name); taxonomy regenerated (107 concepts); SEL-PAIRS -> 11 pairs with pairs[] + pair_reasons{} (status ratified-11-pairs); prose updated. SITE: Selections tab renders each pair as TWO animations side by side with the reason (new pairs flagged); the Collection piece modal shows a 'Deliberate pair' panel with the partner's animation + reason (both pairs for #175), click-through to the partner. Staged to serve dir. Word status unchanged: all 1,536 one-per-pair essence elevens dead (D43b). D44 NOTE (infra): the preview server spawned by launch.json CANNOT follow symlinks into ~/Desktop (macOS permission) — symlinked pgifs/ and gifs/ 404 even though they resolve on disk. Assets must be COPIED into the scratchpad serve dir (handoff rule confirmed). Fixed by copying the 21 pair GIFs + all 512 thumbs as real files, and adding a thumbs/ fallback on every modal/pair . Full-res GIFs for non-pair pieces fall back to their thumbnail. ## D45 — PUBLIC-SITE GAP closed: bitganspuzzlebook.com now carries today's work (2026-09-06) Michael: "I'm looking at the public. That's always supposed to update." CORRECTION: all of the 2026-08-31 -> 09-06 'website' updates went ONLY to the local scratchpad preview. The public site is a separate Cloudflare Pages + Worker + D1 app (~/Desktop/bitganspuzzlebook) fed by scripts/build_tags.py -> db/tags.sql -> live D1, and scripts/push_findings.py (records/concepts/tags/runs). Neither had been run since before this work began. FIXED TODAY: (1) duplicate-concept bug — my new 'deliberate_pair' ledger type was being auto-wrapped as a twin of the designed concept (7 members); folded the alias into the designed concept -> ONE 'Deliberate pair' concept, 21 members, 11 pairs. (2) tags.sql regenerated (214 tags, 5019 assignments) and applied to live D1. (3) push_findings: today's records (22 deliberate_pair w/ partner+reason, 4 HiddenType misfiles, 11 visible-creature, 3 global-Fibonacci extras, {32,95} burn pair, abnormal-HiddenType selection), the updated deliberate_pair concept, 23 negative sweep runs (D29-D43b), and Pindar's 2022-01-01 Discord quote as intel. (4) pages/index.html: the 11 pairs embedded; selecting the 'Deliberate pair' tag renders every pair as two animations side by side with the reason (new pairs flagged); each paired piece's modal shows a 'Deliberate pair' panel with the partner + reason. Deployed via wrangler pages deploy (project bitganspuzzlebook). (5) update_all.sh now ends with steps 7/8: regenerate+apply live tags, deploy Pages — so 'update the website' means the PUBLIC site from now on. Findings still require an explicit push_findings.py payload. ## D46 — 2026-09-08 — BIP-39 color-word essences: complete set, swept dead FACT: the BIP-39 list holds exactly ten plain color words (black, blue, brown, gold, green, olive, orange, pink, silver, yellow) and every one is an essence exactly once (5 robo, 3 ghost, 1 solstice #262, 1 spooky #196; none in skull). Seven color-ish essences also appear once each (color, bright, pumpkin, avocado, coffee, rose, cinnamon). Rainbow is the most-glitched trait value in the collection (5 corruptions: #48 reverse, #58 delete, #66 substitute, #170/#175 'Raibow'); 'Rainbot' = Rainbow with w->t, the only truncated robo stem, and its cut lands on the BIP word 'rain' = essence of the first Rainbot #134 (robo essences are hand-chosen). Recorded as observation, not a word source. SWEEP colors11 (plant-validated, 4 workers): 94,106 orderings / 12.0M seeds — ten colors + each color-ish 11th word (first/last/inline) under gan/mint/alpha/spectrum/drop orders fwd+rev, plus every eleven from the 17 color-ish words in those orders. Plant seen (w4). NO HIT. Registry +94,106 keys. Essence register stays dead (D36 unchanged). ## D47 — 2026-09-09 — Public site brought current: theories 151-173, principles refreshed, five new tag families Michael: "update our puzzlebook website with all the theories we've tried since the last update... clearly and concisely... include any tags we haven't incorporated." THEORIES: 23 new plain-English cards (#151-#173) on bitganspuzzlebook.com covering D29-D46: HiddenType decode (151), abnormal-HT eleven (152), four misfiles (153), global number-class (154), bookends (155), passphrase (156), fib-onset ordering (157), Review remakes/mirrors (158), 32 board (159), untested-methods swarm (160), filename words (161), no-guess rule (162), creature eleven (163), #300 (164), painted marks (165), PINDAR-CONFIRMED corrupt eleven (166, OPEN), verbs-on-numbers (167), undo-pointers (168), teach-the-method (169), eleven deliberate pairs (170), slot-order (171, OPEN), Rainbow/Rainbot (172), colour words (173). #145 updated to point at 151. Each card = picked / tested / happened / status, with piece animations where a card is about specific pieces. PRINCIPLES: all 104 (24 selections, 47 registers, 33 orderings) re-pushed with current status, plain prose, Pindar rationale and members. Live had every register/ordering as 'untested' and 22 stale selections since the first migration; now register/ordering status derives from methods.json 'tested' (tested/partial/untested/dead). TAGS (verified data facts, generated by build_tags.py so they persist): hiddentype_class: (decoded letter), numclass_gan: and numclass_mint: (global number class, D31), filename_word: (85 Arweave filename words, D35), creature: (visible creature ONLY where the word appears in the piece's own name/type/gang/essence or Michael adjudicated it, D38), coburned (10 same-transaction burns incl. {32,95}). 241 tags / 6,238 assignments. INFRA: push_findings.py now accepts 'principles' and 'hypotheses' keys (INSERT OR REPLACE by id). ## D48 — 2026-09-09 — Astra briefing published + public data bundle Michael: brief a friend's GPT-6 Astra with everything tried, how to use the site, where the files are (ours + Pindar's Arweave sheet), the high-quality multi-step ideas (glitches-teach-the-method), a fact-check/tag-audit mandate, all Pindar lore incl. off-site, and a section on what Astra can add. PUBLISHED: bitganspuzzlebook.com/brief/ (designed page; nine parts: puzzle in one page, Astra capabilities, resource map, verified facts, dead space, the ideas, mandates, rules of engagement, glossary) + /brief/BRIEF.md (paste version). DATA BUNDLE: bitganspuzzlebook.com/data/ — 61 files / 35 MB with manifest.json: our spreadsheet (CSV+JSON), deck_512, Pindar's Arweave properties CSV, chain traits + mismatches, DECISIONS, SOLUTION_SPACE, ledger, dictionary, taxonomy, methods, pb_content, pairs, interpretation matrix, dead_sets.txt (146,474 sets), harness source + bip39, Michael's visual notes, burn/forge data, token_index, Arweave tx ids, the long-form syntheses and the self-contained HTML explainers. NOTE: this makes the working files (incl. Michael's notes and the dead-set registry) publicly downloadable; nothing in them is secret, but Michael can pull any file by deleting it from pages/data/ and redeploying. ASTRA RESEARCH (sources in the brief): released 2026-09-03; 1.05M context; text+image in; computer/browser use (OSWorld 2.0 72.6%), hosted shell, code interpreter, web/file search, MCP; effort up to max; >272K tokens billed 2x. Value for us: whole dataset in one context, bulk frame-level vision over all 512 GIFs, free browsing of bitgans.com / Arweave / OpenSea / Etherscan / Discord, independent re-derivation of every claim. ## D49 — 2026-09-13 — Tag schema reorganised by PROVENANCE; evidence shown on the grid; creature narrowed Michael: distinguish validated traits/metadata from chain-only data and chain discrepancies from visual descriptors you can only see by watching; no "glitch candidate" group (refile everything); every tag a plain, intuitive, non-confusable label; when filtering by a tag the grid must show the relevant data without opening each piece. SCHEMA (tag_schema.py, single source of truth; build_tags.py reads it): seven groups by where the fact comes from — Drop · Published traits · Chain-only & discrepancies (chain-only facts / "Sheet and chain disagree" / name defects) · File facts (frame counts & timing / re-saves & lineage / filenames) · Visual descriptors (motion / figure / eyes / background / marks / colour) · Number rules (number class / hidden-type code / prime rule) · Relations & lore (pairs / people / story / collector notes). Second axis = verification chip on every tag: data | eye | candidate | note. RULINGS: (1) "Glitch candidates" and "Observations" retired as groups; all 72 tags refiled. (2) Sheet-vs-chain disagreement is ONE direction-neutral sub-group (Michael: neither side is canonical); the five verb tags describe the kind of difference; which side is the odd one out is stated per piece from the rest of the collection, and for the substitutions (both values real words) it is a judgement from title + art, shown as such. (3) The eleven machine- labelled "auto:" tags renamed/refiled; two duplicates merged (quantum group; CORRUPTED-type = Abstract). (4) Visible creature limited to SPECIFIC creatures {whale, spirit, heart, warrior, shark, blob, cat, dog, wolf, rat, sprite, clown, bird}, anchored in the piece's own name/type/essence or Michael's eye (D38); species labels (skull, ghost, robot, monster, invader…) excluded → 15 pieces. (5) Spreadsheet/CSV exports carry a second header row naming the group. EVIDENCE: piece_tags gained an `evidence` column (migration 0002); build_tags writes a per-piece line for every assignment (sheet value · chain value for disagreements; measured vs standard frame count; filename; partner + reason for pairs; the recorded description otherwise). The Collection grid shows the active tag's definition and each card's evidence lines; the piece page groups tags by the seven groups with evidence inline. NEW LEDGER FACTS (computed from data): wrong_speed on {155,446,467,468,469,512} (all-0 ms where the drop holds the first frame 100 ms) and #228 (all 50 ms); timing_hold on {146,479,480} (100 ms hold on a frame inside the loop); double_mint on #390 (mint slots 519 and 583 both 'Frozen ghostGAN (90/144)', verified in chain_raw_by_mintindex). CORRECTION: the old note "ten pieces run at 0 ms" conflated the skull drop's normal 0 ms timing with the anomaly. Review artifact: lore_catalog_runs/tag_schema_review.html (build_tag_review.py). ## D50 — 2026-09-13 — Every non-visual tag re-derived from PRIMARY sources; gaps found and fixed Michael: revalidate every item in every category except the visual ones; nothing missing, nothing tagged wrong. METHOD: primary_facts.py recomputes chain/file/number facts from Pindar's Arweave sheet, the raw chain traits, the GIF bytes (frames, delays, md5, XMP ids with UUID-v1 timestamps), token_index and the burn transfers — never from the ledger. build_tags.py now takes those tags from primary_facts, so they cannot drift again. verify_d49.py (first pass, 165 checks) and verify_d50.py (second pass) are the audit trail. FOUND AND FIXED (now tagged, with evidence, and recorded in the ledger): - #66 is burnt on chain but the deck said not → burnt now 30 (from burn transfers). Co-burned = same tx {32,95}, {11,20,49,74} or same block {205-208}; evidence says which. - Sheet-vs-chain recomputed on the complete chain pull: 29 pieces differ (the old diff had 20; it was run on a pull that returned no attributes for many tokens). NEW chain-side oddities never tagged: #4 'SkullGAN' (capital S), #88 GANg 'skulGAN', #97 title suffix cut to 'Slender Glitch', #100 'Primoridal' (same move as #16), #124 'skullGANs' plural, #66 stray backtick after the name, #39 GANg cyclopsGAN vs skullGAN, #509 GANg roboGAN vs skullGAN, #33/#290/#315 Dimensions/Eyes differ, and the systematic Twin→Twins SubType rename on all five sheet-'Twin' pieces (#47, #59, #73, #121, #124). Tags are direction-neutral (D49); evidence names the rarer side. - #174 double space missing from the tag (now 7). #383's misspelled trait KEY 'Subtype' tagged. - Chain-only facts added: no Essence trait on chain (every skull and spooky piece, 144); Name only on chain (#268, #379, #401); no chain record in our data (14 Review pieces = exactly the ones the sheet numbers in binary/hex: #457-466, #497-500) — also tagged 'Numbered in another base'. - Description text differs from drop-mates: 7 (was 1): #130 '1024 animations', the portrait/tribute pieces #147/#228/#246/#264/#282, and the finale #512. - Typos split by where they live: name (#16, #100, #123, #124), trait value (#73 Gltich sheet, #88 skulGAN chain, #170 Raibow BOTH sides, #175 Raibow sheet), filename (#9, #73 gltich; #128 raindow). - Timing re-measured per frame: later-drop norm is EVERY frame 100 ms (skull 0 ms). Wrong speed = 11 (ten all-0 ms: #155, #176, #180, #182, #190, #446, #467, #468, #469, #512; #228 all 50 ms). One held frame in an otherwise 0 ms file: #479 f27, #480 f24, #146 f0. The deck's deduplicated delay column had hidden #176/#180/#182/#190. - Essence not a plain word = {#191 err0r, #363 TRUE, #473 FALSE} (case-sensitive), moved to Published traits. - Number-class vocabulary unified; mint numbers verified against token_index (keys are piece numbers). VERIFIED AS RECORDED: drops; every published trait value; essence order (LIS: #21, #57 skull; #259 bless, #268 bird solstice); forge pairs = the 30 burnt; prime re-save batch (Photoshop 22.3, 27 Aug 2021 = exactly the primes 7..127 minus #13/#89); #300 the only file with an OriginalDocumentID and a real Photoshop document id; hidden-type letters and the four misfiles; palindrome F set; number classes. OPEN FOR MICHAEL: (1) SEL-CHAIN-CORRUPT-11 was ratified on the incomplete diff; the complete diff shows chain-side oddities on #4, #88, #97, #100 (+#39/#509 as judgement calls) beyond the eleven — the selection is left unchanged and flagged. (2) #142 sits at robo position 13 (prime AND Fibonacci) yet is in the prime-position set of 15; the rule as written gives 14. (3) #193's HiddenType F at spooky position 0: palindrome or fifth misfile? (4) Re-save sessions (86) are not a simple save-day rule; timestamps are now decoded for every file so the rule can be re-stated. NOT RE-DERIVABLE (lore/judgement, left as recorded): name_bip_proper, error names, tributes, promise pieces, quantum group/capstone, essence_meta, anachronism, negative frames (byte detector not re-run). ## D51 — 2026-09-13 — Corrupt chain renamed and widened; #193 laid out; save-session rule restated; lore validated Michael: if the corrupt chain is more than eleven, rename it and add the others; more information on #193; restate the save-session rule from the data; validate the lore tags and say what cannot be validated. CORRUPT CHAIN → SEL-SHEET-CHAIN-CORRUPT "Sheet-vs-chain corruptions (26)": every piece where the sheet and the chain give different values (complete chain pull, D50), excluding the systematic Twin→Twins rename; 14 Review pieces have no chain record. Members [4,16,17,33,39,44,48,58,62,66,73,81,88,97,100,123,124,128,173,175,281,290,309,315,383,509]. SEL-CHAIN-CORRUPT-11 kept as superseded. Site: union tag "Sheet and chain disagree on a value (all kinds)" + the kind tags. #193 FACTS (no ruling): the other 15 spooky letters follow the letter scheme exactly on their octal position. #193 is position 0 (0o0). Robo position 0 (#129 RoboboR) is also F, and every robo F sits on a binary-palindrome position AND a palindrome name; #193's name is not a palindrome, and the other spooky binary-palindrome positions (#199 0110, #202 1001, #208 1111) are not F. 0 is Fibonacci, so the scheme would give A. Journal 002 defines a–e only and never mentions F. Tag renamed "Hidden-type letter F" with this evidence; not counted as a misfile. SAVE SESSIONS restated from data (save_sessions.json): 39 sessions at a 3-hour gap. Two objective tags replace the old 86: "Exported with a point-release Photoshop build" (77: 22.3 prime batch 26; 22.5 five robos; 23.0 solstice 18 incl. #300 alone on the solstice; 23.1 ghost automated batch 16 + final review session 12) and "Saved alone or as a pair, outside any batch" (10: #1, #121, #152, #228, #300, #448, #467, #468, #469, #504). The old 86 = these two sets minus overlap plus a few judgement members; the old rule was never written down as a rule. LORE VALIDATED: promise pieces #32/#119 confirmed verbatim in journal page 002 (chip → data). Quantum group: 19 by vocabulary in name/traits (qubit, Entangled, Quantum, Spin); #133 and #501–#504 by the collector's eye only. Named-piece seed-word tag now data-derived (18; adds #209 Solar, #300 Lunar). Error name #404 'Not Found' at number 404, essence 'crash': confirmed. Day/night #209/#300 confirmed. #512 capstone confirmed. Tributes: #228/#246/#264/#282 confirmed by the on-chain description; #445/#446/#447/#448/#482 by the name alone (external knowledge; #482 is a joke name); #210/ #211 (solvers), #1/#2/#3 and #130 NOT validated by any data — evidence lines say so. 'Anachronism' (#465/#466) retired: they are January-2022 Review pieces with spooky-style names; the facts are tagged as odd numbering and wrong base. Essence-meta (38) is a curated vocabulary and stays a note. ## D52 — 2026-09-13 — #193 ruled: F is a robo-only class, so #193 is the FIFTH hidden-type misfile Michael: "spookyGANs don't have palindromes. That was only for the Robo Drop." Consequence: the palindrome class F is defined only in the robo drop (eight robots, each on a binary-palindrome position with a palindrome name). #193 (spooky position 0, letter F, scheme gives A) is therefore a misfile, joining #32, #119, #161, #187 → SEL-HIDDENTYPE-MISFILE = 5. Tags: "Hidden-type letter is wrong on purpose" (5), "Hidden-type F: robo palindrome" (8). Brief, taxonomy, prose, theory card 153 updated. ## D53 — 2026-09-13 — "No chain record" was our mistake: all 512 pieces have chain traits Michael asked whether the 14 missing chain records were a data gap on our side. They were: chain_raw_by_mintindex.json holds them at mint slots 597-608 and 642-646 with full traits, but the labelling step keyed records by the '(n/512)' counter in the name and left the binary/hex-numbered Review pieces (#457-466, #497-500) as 'gap'. primary_facts.py now maps chain records by mint slot via token_index.json (fallback: the label). Result: 512/512 chain records; the fourteen match the sheet exactly on every field (no new discrepancies; sheet-vs-chain stays 26). Tag 'No chain record in our data' retired; 'Numbered in another base' (14) stays as the naming fact. Brief, selection text and theory 174 corrected. ## D54 — 2026-09-13 — File facts fully re-derived; number-class descriptions explain Pindar's precedence FILE FACTS: all eleven file tags now come from the files themselves. Newly re-derived this pass: (1) negative frames — a calibrated scan of every frame of all 512 GIFs (frame is negative when its colour inverse is far closer to the previous frame than the frame itself; verified on #231/#479 positive, #5 clean) finds exactly {231,256,274,479}: the trio share an identical 11-frame list (37,38,43,44,47,48,50-54), #479 has 13. (2) Filename words — recomputed from the archive filenames by a stated rule (split on separators and digits; keep seed-list words that are not a trait value anywhere in the collection and not in the piece's own metadata); multi-valued per piece (86 pieces: seasons 29/26/26/20, update 9, ancient 7, smooth 3, elder 1). 'whale'/'spirit' dropped (they are the piece's own GANg). SOURCE CAVEAT recorded on the tag: the filenames come from the collector's downloaded archive (339 files); 64 are confirmed verbatim on Pindar's ArDrive listing (arfs_names.json); 25 pieces have no filename on record (#57, most robos 134-192 range, etc.). (3) Filename typos {9,73,128} from the same names. Already re-derived in D50: frame count, wrong speed, held frame, byte identity, point-release build, solo session, prime batch, PSD lineage. NUMBER CLASSES (Michael: "shouldn't each be half the collection?"): No — Pindar's journal 002 files each number ONCE in a fixed order (Fibonacci, remaining primes, squares, remaining odd, remaining even), so 'even (remaining)' = 243 not 256 (2, 8, 34, 144 are Fibonacci; 4, 16, 36... are squares). Relabelled to his 'remaining' wording; every number-class description, the sidebar group text, the taxonomy definition and the brief now explain the precedence and why it matters (B = the class that glitches, A = the protected set). ## D55 — 2026-09-13 — Everything confirmed against Arweave itself Michael: "go out on Arweave and confirm everything." Done from Pindar's ArDrive (drive cbbca83e-8cf4-44c0-b193-f886f1352cb6), enumerated with Arweave GraphQL (241 file entities in 4 folders: bitgans_on_arweave, bitgans_ghostsinthemachine, bitgans_act1review, noungans) and downloaded through gateways that serve large data in ranged chunks (arweave.fllstck.dev, permagate.io; arweave.net times out on the archives and ar-io.dev wants payment). RESULTS: (1) The properties sheet on Arweave (bitgans_act1.csv, tx UwiYOGcO…) is byte-identical to our bitgan_properties.csv (md5 fb733332…). (2) All 512 GIFs are byte-identical to our local copies: 128 skulls from 01_skullgans.zip, 64 robos from 02_robogans.zip, 16 spookies from 03_specialgans_set1.zip, 92 solstice from 04_solgans.zip, 144 ghosts (zip and loose files agree), 68 review (zip and loose files agree). The only Arweave GIFs with no local match are the two noun-test files (#0's blue and red studies), outside the 512. (3) Filenames: all 339 collector-folder names match Pindar's archive names exactly; the archives supply names for the other 173 (robos '05_000101.gif' style, spookies '0o0.gif', ghosts '423_ghostgan', review '451_reprise01' etc.). The deliberate-pair {14,75} is a single byte sequence stored under two names on Arweave (014_rainbow_skull.gif and 075_flaming_glitch.gif). (4) Arweave lastModifiedDate of the archives: skull/robo/spooky/ solstice zips 1 Jan 2022 (the day of Pindar's 'intentional errors' Discord message and journal pages 014/015), ghost and review zips 31 Jan 2022, the sheet 31 Jan 2022, podGAN montage 15 Feb 2022. DATA: lore_catalog_runs/ardrive_entities.json, ardrive_meta.json, ar_hash_243.json, ar_zips.json, arweave_filenames.json; the zips are kept in lore_catalog_runs/arweave_zips/ (~3.4 GB, not in the bundle). primary_facts.FILENAME now comes from Arweave; tag texts updated; the filename-word and filename-typo sets are unchanged (robo/ghost/review names carry no words). ## D56 — 2026-09-13 — What changed on 31 January 2022 (and the timeline around the clue) Michael: identify what changed on 31 Jan 2022; differences between that file and the chain may be the place to look. WHAT 31 JAN WAS: the publication of the complete Act 1 archive. ArDrive: the properties sheet saved 19:42 UTC and uploaded 19:51; 05_ghostgans.zip and 06_act1review.zip uploaded the same day; journal page 015 updated with every archive link and archived on Arweave at 20:12 (page:timestamp 1643659557). The four earlier drop archives went up on 1 Jan 2022 (the day of the 'intentional errors' Discord reply). No earlier revision of the sheet exists on ArDrive, so it cannot be diffed against a prior version. WHAT DID NOT CHANGE: no GIF in the 31 Jan archives was edited after its Photoshop export (zip entry mtimes equal the XMP save times to the minute, UTC-6, for all 212 ghost/review files; only #467-469 were re-touched, on 1 Jan 2022 with the earlier drops). The archived GIFs are byte-identical to ours (D55). THE SHEET ITSELF (the 31 Jan artifact): 513 rows, UTF-8 BOM, CRLF, no non-ASCII characters. Structural artifact: the Name cell of every unnamed ghost and review row holds a single space (191 cells), while earlier drops are blank — the ghost/review block was pasted from a different table. Double spaces in 7 titles (6 spookies + Evil Olive) are in the sheet as well as on chain. TRUE/FALSE are the only non-lowercase essences. Level = 1 everywhere except #0 'infinite'. SHEET vs CHAIN, placed in time (Blockscout transfer histories for all 512; first transfer out of Pindar's wallet = effective mint): every skull-drop disagreement's chain record predates the sheet by five months (first transfers 28-31 Aug 2021), and Krum saw #58's missing SubType on 1 Jan 2022 — so the chain-side oddities (Primoridal ×2, wobniaR, SkullGAN, skulGAN, cut suffixes, the backtick) were on chain before the sheet was written, and the sheet 'cleans' them; the sheet-side oddities (Gltich #73, Raibow #170/#175, Cronenber #123, RainbowShimmer #128, Space-clops #383, the Twins rename, 'Ancient' #44, skullGAN for #39/#509, and the three numeric differences) are what the 31 Jan sheet put in. Pindar's own journal 003 calls #39 'the cyclopsGAN' → the sheet is the odd side for #39. RELEASE TIMELINE (new): ghost drop first transfers 1-3 Feb 2022 — between the sheet and the clue; #446/#448 (1 Feb), #447 (4 Feb); the prize piece #445 moved into the mystery wallet 0x18f8…478d on 5 Feb 2022 11:12 UTC (clue day); #453 on 14 Aug 2022; the rest of the Review drop (incl. all 14 odd-numbered pieces and #512) first transferred 24-28 Aug 2022, bracketing Pindar's 25 Aug 'looking for glitches all wrong' tweet. Per drop: skull 28 Aug-7 Sep 2021, robo 28 Sep-6 Oct 2021, spooky 16-29 Oct 2021, solstice 21 Dec 2021-4 Jan 2022. Recorded as the chain-only value tag "First left Pindar's wallet: " with the exact time as evidence. Data: lore_catalog_runs/transfers.json, first_transfers.json, zip_mtimes.json, journal/*.txt (raw pages), arweave_archived_page015.html. INTEL added to the site: 'These are not Hidden Typos, but part of the puzzle that will only be solvable after some more drops' (hidden-types page); 'Only 1 clue will ever be given…' (5 Feb post); journal 004 on the prime glitch's Fibonacci timing (0.5/0.3/0.2/0.1/0.1 s); the 31 Jan timeline; journal 003 on #39. ## D57 — 2026-09-13 — Which side is wrong, decided by three witnesses; the two candidate lists Michael: Pindar deliberately fixed errors in the sheet that persist on chain, and the sheet introduced others — are those the ones we were looking at, and aren't both good candidate lists? METHOD: for every sheet-vs-chain disagreement, three witnesses decide which side is the odd one: (1) usage — is the value used anywhere else in the collection; (2) the Arweave FILENAME, written by Pindar at export (Aug 2021–Jan 2022, i.e. between the chain record and the sheet) — when it carries one side's word it is a tie-breaker; (3) Pindar's journal (003 names #39 'the cyclopsGAN'). Encoded in primary_facts.witness(); shown per piece as evidence. RESULT (see the tag set 'Sheet corrects a chain error' / 'Sheet introduces the error' / 'Undecidable'): - CHAIN WRONG, SHEET CORRECTS: the chain record predates the sheet by months and the sheet holds the value the rest of the collection (and the filename) uses. This is the set Michael calls 'the errors Pindar fixed in the sheet'. - SHEET WRONG, SHEET INTRODUCES: the sheet holds the odd value against chain + filename (Cronenber #123, RainbowShimmer #128, Raibow #175, skullGAN for cyclops #39, Ghost for #17 whose filename says glitch). #73 'Gltich' is in the sheet AND the filename (Aug 2021), so it predates the sheet; #170 'Raibow' is on both sides. - UNDECIDABLE: both values real, filename silent: #173, #281, #509 and the numeric/case differences #33, #290, #315, #383. - CONVENTION: Twin→Twins (#47, #59, #121; also on #73, #124). Overlap with the old eleven {16,17,44,48,58,62,66,81,173,281,309}: 8 of the chain-wrong set were in it; it also held #17 (filename now says the sheet is odd), #173 and #281 (undecidable); it lacked #4, #88, #97, #100, #124. SWEEP (local, plant-validated, authorized standing): every eleven inside the chain-wrong set, and every eleven inside chain-wrong ∪ sheet-wrong, essence register, mint/gan/reversed orders (sweeps/witness11). Result recorded when done. ## D58 — 2026-09-14 — Pindar's filename layer is his private construction code (verified facts) From the Arweave archives' filenames (D55): (1) skull 's_' prefix marks the squares that take letter c: #4, #9, #16, #25, #36, #49, #64, #81, #100 — every square except #1 (Fibonacci takes precedence, letter a) and #121 (11², letter c, file '121_twin_smoke.gif', the one unmarked square). (2) 'smooth' appears only on #5, #13, #89 — the Fibonacci primes above 3, the primes protected from the glitch rule. (3) 'Primordial' is written as 'ancient' in 8 skull filenames and as 'elder' once — on #16, one of the two 'Primoridal' chain errors. Both are seed words; Primordial is not. (4) Solstice filenames carry the solved puzzle's mechanism: '1221' (the solstice date), a season code 04–10/44 that tracks the chain Date trait by month, and a running number. (5) Ghost filenames are bare ('N_ghostgan'). (6) Review filenames group the drop: reprise00…13 with NO reprise07, plus repriseaa, reprisebs, reprisehighdef, repriseline(s), reprisetest, skullall and the four tribute names. The 14 binary/hex-numbered pieces are exactly reprise04/05/06/08 + repriseaa — the reprises of the robo, crypto and spooky pieces, numbered the way their originals were. Recorded as facts; interpretation in the reading given to Michael the same day. ## D59 — 2026-09-14 — Michael: the substitutions echo visual elements the data cannot see Ruling: every sheet-vs-chain SUBSTITUTION (two different real words for one trait) echoes a visual element of the piece — the substituted word describes what the art shows; the other side keeps the formula label. Applies to #17 (Ghost→Glitch), #39 (skullGAN→cyclopsGAN), #44 (Ancient→Primordial), #66 (Rainbow→Flaming), #173 (Ghost→Smoke), #281 (glitchGAN→bitGAN), #509 (skullGAN→roboGAN). They are therefore NOT errors on either side and leave the 'Sheet corrects a chain error' set, which becomes the letter-level errors only: {4,16,48,58,62,73,81,88,97,100,124,309} = 12 (#58 displacement and #62 swap included). 'Sheet introduces the error' = {123,128,175}. Undecided = the numeric/case differences {33,290,315,383}. New tag 'Substituted word describes what the art shows' (eye, 7). READING: the substitution category teaches that the picture is the authority over the metadata — a label that disagrees with the field is telling you what the piece IS. Five of the seven visible truths are seed words or cut to one: robot (#509), bit (#281), smoke (#173), flame/fire (#66), ancient (#44 via the filename); glitch (#17) and cyclops (#39) are not. D59 addendum: #66 counts in BOTH sets (its SubType Rainbow→Flaming echoes the art; its title backtick is a chain error), so 'Sheet corrects a chain error' = 13 and 'Substituted word describes what the art shows' = 7. New VISUAL tag at Michael's request: 'Name describes a visual element' (Visual › Figure, eye) — a word in the name or traits that names something visible that the data cannot know; wider than 'Visible creature'. Seeded with the seven substitution pieces; Michael will adjudicate all 512. D57 RESULT (2026-09-14): sweeps/witness11 complete — 151,164 orderings / 19.35M seeds (every eleven inside the 14 chain-error pieces, and every eleven inside the 19 = chain-error ∪ sheet-error; essence register; mint, gan, both reversed; 128 twelfth words; 33 paths). Plant seen (w1). NO HIT. Registry +151,164. The essence read on the artist- confirmed selection is closed, as the Solstice precedent predicted; the word channel for these pieces is still open. ## D60 — 2026-09-14 — Stacking the discrepancy categories (Michael's question) Michael: Pindar might stack glitch types; can the other discrepancy categories fill the six words the substitutions leave, as one block or two, and are the discrepancies spread across drops? CORRECTION: 'bit' and 'five' are not on the wordlist, so the substitution block yields FOUR words (robot #509, smoke #173, flame/fire #66, ancient #44), not five. DISTRIBUTION by drop of the 30 discrepancy pieces: skull 21, robo 3 (#170, #173, #175), solstice 2 (#281, #290), ghost 3 (#309, #315, #383), review 1 (#509), spooky 0 — the spooky drop's oddities are name defects only (double spaces ×6, 'Gan' #205, '8bit' #194, F on #193). So a one-per-drop stack cannot be built from discrepancies alone. WORDS the discrepancies point at, one stated step each (table in the session): substitution → the visible truth (robot, smoke, flame, ancient); transposed word → its filename translation (elder #16, ancient #100, twin #73); reversed/deleted/space-dropped Rainbow → cut to 'rain' (#48, #58, #128, #170, #175); skullGAN with a letter added, dropped or recased → 'skull' (#4, #62, #88, #124); suffix cut → the residue (ghost #81, slender #97, frost #309); capital changed → 'space' (#383); Twin/Twins → 'twin' (#47, #59, #121). Glitch (#17), cyclops (#39), cronenberg (#123) and the number changes (#33, #290, #315) yield no seed word. Distinct pool = 12 words: robot smoke flame ancient elder twin rain skull ghost slender frost space. SWEEP sweeps/stack12: every 11 of the 12 words, flame or fire, every choice of piece for the repeated words (which sets the mint order), mint and piece-number order forward and reversed: 2,656 orderings / 340k seeds, plant-validated. D60 RESULT: sweeps/stack12 HALTED by Michael after 4 of 27 chunks (no hit in what ran; not resumed). STANDING RULE (2026-09-14): no sweep without a stated three-part hypothesis — selection, word derivation, ordering — agreed first. ## D61 — 2026-09-14 — The moves as instructions, both directions; cipher-history parallels; four statable hypotheses Michael: repeats are valid; the chain error may teach how to find/transform words; both sheet→chain and chain→sheet may be valid lessons; think creatively but simply, as Pindar might; look at cryptographic history. WORDLIST TEST OF THE MOVES (which moves carry a seed word to a seed word): reverse 5 pairs (edit/tide, evil/live, liar/rail, loop/pool, net/ten); adjacent swap 6 pairs (act/cat, add/dad, alter/later, door/odor, sign/sing, there/three); drop one letter 198 pairs (brain/rain, cloud/loud, crash/cash, exist/exit, army/arm…). The naga names DEMONSTRATE the reversal move on words: Evil Olive → evil + olive (both seed words), Flow Wolf → wolf, Dog God → dog, Taco Cat → cat. CIPHER HISTORY: Kryptos (Sanborn's deliberate misspellings IQLUSION/UNDERGRUUND/DESPARATLY, now argued to encode the K4 key); null ciphers (errors mark which positions count; 18th-c. log-table errors); Bacon's bilateral (case/typeface as binary — cf. robo binary positions and #4's capital); Beale/Ottendorf book ciphers (numbers index a public text — the BIP-39 list is the book). FOUR COMPLETE HYPOTHESES (none run): (1) UNDO AND CUT — selection = the 11 malformed chain labels the sheet corrects (set B {4,16,48,58,62,81,88,97,100,124,309} or A {…66 instead of 309}); word = undo the error, restored word cut to its on-list stem, repeats allowed (skull elder rain rain skull ghost|silver skull slender ancient skull ghost|frost); order = mint. (2) NULL CIPHER — same selection; the error only marks the piece; word from one unpublished column (filename word, hidden-type class name); mint. (3) APPLY THE MOVE — same selection; apply each piece's own move to its restored word until it lands on the list; mint (fails on swaps unless the base word differs). (4) TWO DIRECTIONS — sheet-corrected errors read restored, sheet-introduced errors read corrupted; both sets; mint. Recommendation: (1) first, B, mint — a few dozen phrases. Awaiting Michael's pick per the standing rule. ## D62 — 2026-09-14 — The D61 hypotheses run (inside their stated rules only) Michael: "Run the hypotheses." Each was encoded exactly as stated; a hypothesis that cannot fill eleven slots under its own rule was reported infeasible, not forced. RUN (sweeps/hyp61, 1,008 phrases / 129,024 seeds, plant seen by w2): H1 UNDO-AND-CUT on selections A and B — restored word cut to its on-list stem, repeats allowed; free choices only where a cut suffix reads as the restored token or the residue, and elder|ancient for Primordial; mint and reverse-mint (240 phrases). H2 NULL-CIPHER with the FILENAME word as the unpublished column, selections A and B, mint and reverse (768 phrases). RESULT: NO HIT. INFEASIBLE AS STATED (not run): H2 with the hidden-type class name (only 'square' is a seed word; 7 of 11 pieces have none). H3 apply-the-move (reverse / adjacent swap / drop a letter applied to each piece's restored trait word lands on the list for 0 of 11; applied to the essence, 1 of 11: army→arm). H4 two-directions (of the sheet-introduced spellings only RainbowShimmer cuts to a seed word; Cronenber and Raibow yield none). STATE: selection A/B + mint order stand; the essence, undo-and-cut and filename registers are now all closed on them. The word channel for the artist-confirmed errors is still not identified. ## D63 — 2026-09-14 — Site, brief and spreadsheet brought current with D53–D62 Theory cards 175–181 (Arweave confirmation; 31 Jan 2022; three witnesses; substitutions echo the art; the two candidate elevens and the registers closed on them; the filename layer; the moves as instructions + cipher history). Runs: witness11, hyp61, and the halted stack12 all on the Runs tab. Selections: SEL-CHAIN-ERROR-A and -B added as candidate exact-elevens with registers closed; SEL-SHEET-CHAIN-CORRUPT text now states the four witness piles. Tags: filename-layer facts added under File facts › Filenames ('s_' square mark, the unmarked square, 'smooth', Primordial written as ancient/elder, Review filename group); the disagreement tags' descriptions rewritten to name the pieces and the witness rule. Spreadsheet (build_sheet.py + bundle CSV): 15 new columns — Arweave filename, filename words, filename mark, review group, sheet-vs-chain difference and verdict, chain-error eleven (A/B), first transfer and period, Photoshop save/build/ session, negative frames, decoded HiddenType. Brief: Part 4 (witnesses, filename layer), Part 5 (tests on the elevens), Part 6 (6i the label bends toward the picture; 6j the moves as instructions). ## D64 — 2026-09-14 — Site tabs audited and brought current; testing rule published Definitions tab: concept definitions rewritten where stale (resave_sessions → point-release/solo sessions from decoded XMP; hiddentype → five misfiles incl. #193, F robo-only, precedence order; chain_mismatch → 26 pieces and the four witness piles; held_frame/wrong_speed/negative_frames/primes_missing/psd_lineage → D50–D54 measured facts; name_typo split by layer; double_mint and description_anomaly with verified members). Extraction tab: tested notes updated on S2-W1, S2-W5 (now the leading open register: what the corrected label names visually), S2-W16, EC24, EC28, EC38. Ordering tab: S3-O13 (39 sessions), OC28 (first transfers known). Hypotheses tab: 142 superseded → theory 177/179; 166 and 174 closed as history (tested); 171 pointed at A/B; 9 and 4 updated with the decoded code and the exact sheet errors. Brief: Part 8 rule 1 'How a test must be stated before it runs' (selection / word derivation / ordering, each in a sentence Pindar could have written; enumerate inside, never jam); Part 7 asks the solver to state the three rules; data bundle index carries the same paragraph. All 104 principles re-pushed. ## D65 — 2026-09-14 — Hypothesis rules lead every brief; the bookend shape is a standing option Michael: lead the briefs with the rules around hypothesis creation and creativity, and confirm that hypotheses include the shape 'nine or ten glitch pieces with the Pandimensional skullGAN (#0) and the Quantum glitchGAN (#512) as bookends'. RULES NOW STATED FIRST in the audit brief (section 0) and the Astra brief (intro): (1) no brute force without a stated selection + word-derivation + ordering rule, each in a sentence Pindar could have said; enumerate inside, never jam; motivation not count; repeats allowed. (2) The phrase may be 9 or 10 glitches with #0 (universe / Dimensions infinite) and #512 (final / Dimension 11) as bookends, or 10 with one bookend; a hypothesis must say which shape it assumes; Solstice precedent ('begins with sunrise… ends with sunset'); D30's bookend sweeps closed only the essence read. (3) Creativity in Pindar's register: finds are lessons as much as containers; ask what an oddity teaches before what it contains; vanilla field reads are exhausted. (4) Check us, do not trust us. SEL-ANCHORS renamed 'Bookends: #0 universe and #512 final' with members {0, 512, 357} and the shape rule in its text. Memory rule updated to include the bookends. ## D66 — 2026-09-14 — #357 retired as an anchor Michael: "Kill that as an anchor. It's an interesting observation to keep but that is all." #357 (Shy skullGAN, essence 'seed') leaves SEL-ANCHORS, which is now exactly the two pieces that stand outside the numbered run by their traits: #0 (universe, Dimensions infinite) and #512 (final, Dimension 11). The D2–D4 idea 'seed or final as the eleventh slot' is history only. #357 keeps its tags: label contradicts art (does not hide), colour off the drop's rainbow gradient, near-miss invader, essence 'seed'. ## D67 — 2026-09-18 — The Named pieces: what Pindar says, and what the names mean outside the project Michael asked whether 'Moirai' or 'Geometer' carry significance and which other names have external meaning. PINDAR'S OWN TEXT (Pandimensional skullGAN ch. 2): 'a select few would be Named… the key to unlocking the full story'; the first two Named are The Moirai (#1) and the Pandimensional skullGAN (#0): 'The Moirai to reveal that everything was already determined and the Pandimensional skullGAN to reveal that everything has already happened.' The dream-book opens 'Every story has a beginning middle and end. This is the beginning of the end.' Both quotes added to the site's intel. DATA FACTS: 38 Named pieces. #1/#2/#3 (Moirai, Artist, Geometer) have Eyes 6/4/2 and filenames ancient_triplet / charred_ skull / scholar_skull. All eight nagas are palindromes in EVERY field (Name, Type, SubType, and the GANg 'nagaGAN' itself); naga position 0 has essence 'alpha' and position 63 'zero' ('omega' is not a seed word). #404 'Not Found' sits at piece 404 (essence crash); #420 'Four twentyGAN' at 420 (essence chronic). #449 Name 'Bean' reprises #21 Beanie skullGAN, whose essence 'bean' is one of the two genuinely displaced skull essences. #445's filename is 'colborn' (the collector tribute). EXTERNAL MEANINGS (context only; under D36 never a word source): Moirai = the three Fates (Clotho spins, Lachesis measures, Atropos cuts) → a three-headed skull; Geometer = 'earth-measurer', Plato's 'let no one ignorant of geometry enter', God as geometer with the compass; Tenet = centre of the Sator palindrome square; Sol/Luna = sun and moon (chain Date Day/Night); Magic Internet Money = the 2013 bitcoin wizard meme; slenderGAN = Slender Man; Cronenberg = body horror; Testy McTest = the Boaty McBoatface joke. Seed-list check: measure, artist, three, circle, sun, moon, solar, lunar, day, night, alpha, zero are on the list; fate, thread, destiny, compass, omega, tenet are not. READING: the lore supports the bookend shape (#0 and #1 are the first Named, framing 'already determined / already happened') and that the answer was fixed in the collection before the clue. No hypothesis is proposed from this. ## D68 — 2026-09-19 — New tag category 'Essence & wordlist findings'; Published traits holds only what Pindar printed Michael: findings were showing under 'Named' in Published traits that have nothing to do with the official Named data. CAUSE: six tags in the Published category had no sub-group, and the sidebar rendered un-grouped tags after the last heading ('Named'), so they looked like Named tags. FIX: (1) new category 'Essence & wordlist findings', placed directly above Published traits, for findings made by reading the published words against the BIP-39 wordlist and Pindar's own index column — sub-groups: Pindar's index column (has his BIP-39 index; index decodes a non-word essence; word out of its home block), Essence order (breaks the alphabetical run; swapped with another piece's), Essence form (not a plain word), Word sets (colour word), Names (Named piece with a seed word in its name — moved from 'Name defects on the chain', where it never belonged). 'Word out of its home block' moved here from Number rules › Wordlist index (sub-group retired). (2) Published traits now holds ONLY what Pindar printed: GANg, Type, SubType, Eyes, Dimensions, Named. (3) Sidebar renderer: a tag with no sub-group renders first in its category, never under the previous heading. Piece page and CSV export carry the new group. Removed the 'animate all' checkbox the same day (cards play when scrolled on screen). ## D69 — Whole-taxonomy vet: banded sub-groups, misfiles moved, duplicates removed (2026-09-19) - **Misfile found.** "Name defects on the chain" was wrong as a heading: leet names, odd numbering, wrong base, double space, error names and 18 of 19 species mismatches read the SAME on the sheet and the chain. They are quirks of the printed name, not chain facts. Moved to the renamed category **Name & essence findings** (key `wordlist`), bands Essences (Order / Form / Word sets) and Names (Spelling and typography / Numbering / Meaning). `essence_meta` moved there from Relations › Story. - **Hidden-type letter** is read off the chain, so its parent tag and its five child tags now sit together under Chain (they were split across Chain and Number rules). Duplicate `hiddentype_palindrome` removed (same eight pieces as the decoded value 'palindrome'; its definition was folded in). - **Bands** (a third level, written "Band › Sub-group"; the sidebar prints the band once): Sheet vs chain › Which side is wrong / Kind of change; Filenames › Added words / Marks / Defects / Review groups; Number class › Of the piece number / Of the mint number; Motion › Frame repeats / Colour events; Figure › Species / Composition / Echoes and one-offs; plus new visual sub-group **Label against art** (name describes a visual element, art contradicts its label, named Glitch but does not glitch). Visual membership untouched — structure only. - Chain-only facts split into Wallet and mint history / Forge burns / Chain metadata. Filename-word family no longer repeats smooth / ancient / elder (each has its own tag). - "Which side is wrong" is a true four-way split of the 26: chain wrong 13, sheet wrong 3, undecided 4, neither-wrong-both-describe-the-art 7 (one overlap, #66). Three more pieces (#47, #59, #121) differ only by Twin→Twins and are counted under that tag; the umbrella label now says 26. - New data tag `name_bip_fused` (6): seed word fused into an invented GAN word (acid, method, ice, twenty, rose, slender). With the free-standing tag (18) that is 23 of the 38 Named pieces. - Observations for Michael, no action taken: Review filename groups reproduce our visual families exactly (reprise13 = Abstract GANs, reprisebs = black background = rainbow prelude, reprise12 = beige, repriseline+repriselines = line background) — Pindar's own grouping; `bg_test` = testGAN = Type Tester (39) exactly; `out_of_order` (#467–469, essence 'random' on #468) has no ledger source text; #0 is not on the site grid, so the #0 bookend has no tag. - Two exact elevens inside the 38 Named: Named with a hidden-type letter (3 skull + 8 palindromes; essence set already in the tested registry) and Named whose GAN word is not its GANg (ten invented GANs + #512; NOT tested; piece order + essence does not admit 'random' as a legal 12th word). No sweep run — awaiting an agreed three-slot hypothesis. ## D70 — hyp62: the eleven Named pieces with an invented GAN word (2026-09-19) Agreed with Michael before running ("go ahead"). Selection: Named pieces whose xxxGAN word is not their GANg (#228, 246, 264, 282, 420, 445, 446, 447, 448, 482, 512). Word: essence. Order: piece number (ends on #512 unaided). Shape: eleven glitches + any checksum-legal twelfth word. Run: 1 ordering × 128 twelfth words × 33 paths, plant seen. **No match.** 'random' is not a legal twelfth word in this order. Open inside this selection: another word register (the invented GAN words give only five seed words). Site card #182, selection SEL-NAMED-INVENTED-GAN, run logged. ## D71 — Beige backgrounds are new to Act 1 Review (collector's ruling, 2026-09-19) #479 and #480 added to "Background style new to Act 1 Review" (now 12 pieces) via the ledger. Both keep the Beige background tag. A dot-colour key was added to the top of the tag sidebar the same day. ## D72 — Site consistency pass (2026-09-19) Category text on the page now comes from tag_schema.py. The Definitions tab follows the schema's band and sub-group order and has a blurb for every sub-group (the old blurbs were keyed to retired family names). The brief's site section was rewritten (current tab names, 182 cards, 48 registers, sidebar grouping, dot key, evidence lines) and gained the Named-pieces work. The audit brief carries a dated note on the regrouping. The /data bundle was stale (ledger, taxonomy, sheet, glitch dictionary, pb_content; spreadsheet from 14 Sep): `refresh_public_data.py` now runs inside update_all.sh (step 7b) and `export_public_sheet.py` rebuilds the 68-column public spreadsheet. tag_schema.py and build_tags.py joined the bundle. ## D73 — The early testGANs, and order breaks in Pindar's own Review records (2026-09-19) - #467–469 were saved 23–24 Oct 2021: eight days after the spooky drop files (15 Oct 2021), ten weeks before any Ghost-drop testGAN (all 32 saved 6 Jan 2022), three months before the rest of the Review drop. They are the first testGANs made. Pindar placed them after the spooky reprise (#465–466), which is where they fall in time. Collector adds: three different eye sets (#467 rare, #468 skull, #469 robo) — in Oct 2021 only skull, robo and spooky eyes existed. - Pindar's reprise group numbers follow drop order (00–03 skull, 04–06 robo, 08 spooky, 09–10 Solstice bitGANs, 11 Ghost monsters, 12, 13). Breaks: reprisetest split (#467–469 / #481–484); reprise11 (#470–472) ahead of 09/10; #477 named reprise11 among 09/10; no reprise07. New data tag `file:review_group_out_of_order` (7). - Mint order breaks: #24, #30, #310, #311, #359, #390 (double mint), #445, #446 (minted last beside #512). New data tag `chain:mint_out_of_order` (8). - Second half of Review (#481–512) runs in save-date order by group: line and skullall 6 Jan, test 8 Jan, highdef 10 Jan, aa 25 Jan, bs 26 Jan, 13 on 30 Jan, #512 last. First half (#449–480) was all saved 29–30 Jan and follows drop order instead. - Michael's theme, recorded as an open idea: "out of sequence" recurs at every level — frames (the standard glitch), letters (Primoridal, Gltich, wobniaR), trait fields (Type/SubType swap), essences (alphabetical breaks, displaced blocks), pieces (Review order, mint order), files (save order). No test run. ## D74 — "Out of sequence" becomes a theme that gathers the specific tags (2026-09-19) Michael's ruling: one larger tag that incorporates every out-of-sequence finding, without losing any specific tag. Built as a new top sidebar group, **Themes across groups**, holding "Out of sequence (all kinds)" (81 pieces) and one roll-up per level: frames 53, pieces 15, essences 8, letters 4, files 4, trait fields 1. The theme is computed from the specific tags in build_tags.py (OOS_PARTS); none of them moved, and each theme card lists which kinds apply. New data tag `file:save_out_of_order` (#129, #147, #159, #203: the only breaks in the two drops that were saved in number order). Correction: #359's essence is 'bright', not 'seed' (that is #357); fixed in the mint-order definition and D73. ## D75 — Four new hypotheses from today's findings, none run (2026-09-19) Cards #183–186, all open, awaiting Michael's go-ahead. (183) the eleven broken essences [21,57,129,131,141,189,191,259,268,363,473], essence as printed, piece order; 'action' (the one unheld index) is a legal twelfth word. (184) eleven Named pieces whose name word is another piece's essence, word = the name word (untested register), order = own number or pointed-at number; #512 is a twelfth pointer to #130. (185) eleven pieces whose number disagrees with mint or make time, essence, mint order; contains 'random' (#468). (186) the glitch as an ordering lesson; needs a frame-replay measurement first. Checksum arithmetic only. Honest note: across the phrases checked today, meaningful twelfth words passed the checksum at about the 1-in-16 chance rate, so a legal 'action' or 'random' is not evidence by itself. ## D76 — hyp63–66: the three new selections run, then alternative orderings (2026-09-19) Agreed with Michael ("try all of these, and if they don't hit think of different orderings"). Cards 183 (broken essences), 184 (Named pointers, name-word register) and 185 (number vs time, plus the universe…random shape) run as stated: 5 phrases, no match. Then 27 further orderings, each a plain rule: piece number, mint slot, date the file was made, the word's wordlist place, date it first left Pindar's wallet, each forwards and backwards, plus three pointed-at orders for card 184. All × 128 legal twelfth words × 33 paths; plant seen in both runs. **No match.** Cards 183–185 set to tested; card 186 stays open with the existing measurement (one frame returns at gaps 5, 3, 2, 1, 1) and one literal reading (a word that returns at places 1, 6, 9, 11, 12) that needs Michael's view. Open inside the selections: corrected words on 183; the Named pieces' own essences on 184. ## D77 — hyp67–68: card 184 with essences, and the orderings the glitch teaches (2026-09-19) Agreed with Michael ("Run option 1. For 186… any ordering/selection methodology you can state simply as a sentence or two, try it out"). (67) Card 184's eleven with the Named pieces' own essences, 13 plain orders: no match. (68) Card 186. Measurement (lore_catalog_runs/fib_return_frames.json, near-match on 64×64 frames): all eleven backwards-Fibonacci pieces return one frame six times at gaps 5, 3, 2, 1, 1; eight runs end on the last frame of the loop; three start elsewhere (#182 frame 19, #190 frame 29, #287 frame 22; 19 and 29 are the robo positions of #148 and #158 — noted, nothing drawn). #291 adds a seventh visit (gap 2), confirming the collector's ruling that it is out. Rules run: order in which the glitch begins (3 orders on the backwards-Fibonacci eleven); Fibonacci walks with repeats, 8 versions, on 13 elevens. 91 orderings × 128 × 33 paths, plant seen. **No match.** Cards 184 and 186 set to tested. Not run: one word returning at places 1, 5, 8, 10, 12 (does not fit "eleven glitches"). ## D78 — Glitch instances, a review of past runs, and the 'solution' block (2026-09-19) Agreed with Michael ("run with this… ID additional interpretations… come up with new hypotheses"). - **Instances.** 34 separate sheet-vs-chain glitches on 29 pieces (lore_catalog_runs/glitch_instances.json; on the data page). Two instance rules with repeats run (one character dropped or added: 11 on 10 pieces, 'author' twice; whole-word glitches outside the species field: 11 on 9 pieces). Measured caution: of ~1,100 one- and two-condition rules, 14 pick exactly eleven. Label registers (Type, SubType) are infeasible: at most 9 of 11 are seed words. - **Wider readings of past recipes** (card 188): plain orders on the two Named elevens; mint-slot word on chain-error elevens A, B and 'chain wrong, not a substitution'; bookend shape (universe + ten + any; universe + nine + final + any) on eight selections; broken essences put back where the index says (alpha twice). 705 phrases. - **New structural fact** (card 189, new data tag `wordlist:essence_solution_block`, 12): the Review essences are scattered over the wordlist except twelve on consecutive words 1655–1666, solution…south; 'solution' is first by wordlist and by piece number; each of the twelve pieces is from a different Review working group. Every eleven of the twelve × plain orders: 96 phrases. The fixed twelve fail the checksum in four plain orders. - All runs: 128 legal twelfth words × 33 paths, plant seen, **no match** (hyp69–71, 801 phrases). - Also noted: the order-the-glitch-begins ordering run as hyp68 had already been covered by the earlier 'fibloc onset-orderings' run; and a name-word register had been tried before on a different eleven (ProperName-11), so card 184's "never tested" applies to that eleven only. - Open cards needing Michael: 190 (TRUE/FALSE as a yes-or-no test; needs his eye on #363 vs #473) and 191 (eleven kinds of glitch, first piece of each; needs his list of kinds). ## D79 — Systematic sweep, the drop rule, the pair chooser, eleven broad kinds (2026-09-19) Michael's instructions: (1) a glitch that drops something may drop the piece; (2) test multiple orderings for every selection × word-rule combination, and the shape ten words + 'final' + random twelfth; (3) TRUE/FALSE may choose which member of each pair; (4) explore one word per working group; (5) distil all glitches into broader categories and check nothing is missing. - **hyp72** (29,901 phrases, 3.83M seeds, 8 workers, plant seen, **no match**): 16 selections × 7 word rules (essence; word at piece number / mint slot / place in own drop, from 1 and from 0) × 12 orders × 2 shapes (eleven + any; ten + final + any, each piece left out in turn). Drop rule: 29 disagreeing pieces − 14 with a drop/add glitch − 4 undecidable = exactly eleven [4,16,17,39,44,48,62,100,173,281,509]; also every eleven of the twelve that keeps #383. Pair chooser: all 2,048 one-of-each-pair choices × essence × 12 orders (repeat kept when #175 is chosen twice). - **Eleven broad kinds** now live as themes (category `theme`, built in build_tags.py THEMES from the specific tags, nothing moved): out of sequence 81, dropped 29, added or doubled 115, swapped 73, reversed 29, written in another code 31, wrong speed or time 47, misspelled 16, hidden in the art 15, colour breaks 30, echoes another piece 31. We did not aim for eleven; 'misspelled' overlaps three others, so a stricter count is ten. Not ratified by Michael. Tags outside every theme are listed in lore_catalog_runs/tags_outside_themes.json (mostly descriptive facts: backgrounds, Review groups, first transfers; glitch-like leftovers: larger/still/unique eyes, abstract class, visually novel, description anomaly, point-release saves). - **hyp73** (card 191 provisional, 74 phrases, no match): first piece of each broad kind = #1, #4, #5, #7, #7, #9, #9, #17, #48, #130, #131. - **Working groups.** Each of the twelve so-words is in a different Review group; the groups without one are the four tributes, reprise01, 03, 05, 08, 09, 10, repriselines, skullall and #512's. No constant position inside the group. Other near-neighbour essence pairs in Review: broken/bronze, call/calm, knee/knife, network/never, reveal/reward, trade/tragic. #473 FALSE sits beside #474 'truth' in reprise09. No rule found yet; nothing run beyond card 189. - Noted: SEL-BROKEN-ESSENCES (card 183) has the same members as the older SEL-SCHEME-EXCEPTIONS, which had been run in January under a different word normalisation. ## D80 — 'solution' as a heading, the block as a key, and the reprise map (2026-09-19) - **hyp75** (card 194, 304 phrases): the eleven so-words under the heading, ordered by Pindar's group numbers, place inside the group, the drop each reprises, his journal-002 number classes (piece number and mint slot), eyes, dimensions, title/Type/SubType A–Z, frame count, Fibonacci walks; both shapes. No match. In piece order nine of twelve so-words are already in wordlist order; soup, sorry, sort are out of place. - **hyp74** (card 195, 7,144 phrases): the block as a transposition key, 12-place [1,2,10,6,3,4,5,8,7,9,11,12] and 11-place [1,9,5,2,3,4,7,6,8,10,11], applied both ways to 17 elevens × 7 word rules × 3 base orders × both shapes. No match. - **hyp76** (card 196, 31 phrases) and a new map: 22 Review pieces share an exact title with an earlier piece, 13 match exactly one (lore_catalog_runs/review_reprise_map.json). #450 'solution' → #411 'key'. Binary Review counters minus the leading 1 → robots: #458 → #130 'quantum', #463 → #135 'ghost', #499 → #139 'echo'. Ten distinct pointed-at essences run as ten + final + any and universe + ten + any. No match. Open: the rule that carries a glitch from the earlier piece onto the Review word (Michael's 'glitches teach, apply to Review'). All with 128 legal twelfth words × 33 paths, plant seen. ## D81 — No mixing: a set of glitches is all data or all visual (collector's ruling, 2026-09-19) Michael: "When we look at Glitches in sets, I think it's either all-chain/data or all visual. No mixing." Standing rule (memory: feedback_no_mixing_data_visual). Every selection must state its side and every member must be known from that side alone. - Site: each broad kind under Themes across groups now has a "from the data" tag and a "by eye" tag; the "all kinds" tag mixes both and is marked as browsing only. Split by side the kinds are nine and nine (data: wrong time, misspelled, echo, dropped, added, out of sequence, swapped, reversed, another code; by eye: swapped, added, out of sequence, dropped, hidden, reversed, colour, echo, wrong speed). - Audit of selections already run. All data: chain-error A, B and 'not a substitution', old chain-corrupt, broken essences, the three Named elevens, number vs time, letter-level typos, title glitches, drop-rule eleven, colour essences, the solution block. All visual: backwards-Fibonacci eleven, creature eleven, mislabel eleven. MIXED: the eleven deliberate pairs (8 known from data; 3 known only by eye: #489/#491 scanline, #130/#133 flash, #165/#207 painted numbers), the halted stack12, and the first card-191 run. - hyp77: card 191 rerun one side at a time in the bookend shape (universe + nine + final + any), 10 phrases, plant seen, no match. ## D82 — The collector's Review Desk answers (2026-09-20) - **Kinds ratified.** All eleven kept; a twelfth added: **Review Drop Natives** (his name for 'looks like nothing in an earlier drop'; tag relabelled). 'Chain description differs' filed under Swapped. Larger eyes, still eyes, unique eyes, Abstract GANs and point-release saves ruled descriptions, not glitches. By side the kinds are now nine (data) and ten (visual). - **TRUE/FALSE.** #363 is a standard ghostGAN for its drop; #473 is a totally new kind (invader background, oval eyes with rectangular NounGAN pupils); #474 'truth' is the same type as #473 and disappears from frames as Ghost types do outside the Ghost drop. Reading recorded: TRUE = true to its drop, FALSE = a new kind. 'Dancing' checked in the file data: the 15 Dancing pieces have the same 55 frames at 100 ms as their drop-mates, so it is not a frame-rate trait. - **#467's eyes** are spooky eyes, the same as Monster spookyGAN #207. The three October 2021 tests wear spooky, skull and robo eyes: the three drops that existed. - **solution → key**: unsure; something extends from #450's head for a few frames and may look like a key. To be checked frame by frame. - **Carrying rule**: he leans to 'the Review word is replaced by the earlier piece's word', not sure; wants to look at the set together. - **Name describes a visual element**: in progress; reframed as finding the outliers (smokes, silvers, whales, beans). - **New standing list**: theories/DEEP-REVIEW-LIST.md, every 9–12 set from the past days, run or not, each to get an in-depth review page with all data beside the animations. - **hyp78** (548 phrases, plant seen, no match): ratified visual kinds as ten + final + any and universe + ten + any; the carrying rule on the so-words (title map; title map plus binary counters; 'key' standing first). ## D84 — The so-GANs are not recoloured or retimed copies (measurement, 2026-09-20) Michael's question: could the so-GANs with no title match be older GANs colour-inverted or changed, or pieces on the same altered timing or colour cycle? Measured on the stored 32×32 frames (22,753 earlier, 4,028 Review): normalised correlation of outlines, brightness, |brightness| (inversion allowed), colour intensity, each also mirrored; per-piece colour-cycle and motion-rhythm series compared with circular correlation; frame counts and delays from the file data. Result: no copy. Best score in the Review drop 0.70 (#494 → #423, #360); #458 → #150 0.64 (as the earlier direct-colour study found); Inverted pieces #479/#502 rise only to 0.45/0.55 with inversion allowed; colour-cycle and rhythm matches at or below the baseline any two pieces reach (median top-1 0.86 / 0.81); #479's held-frame timing and #494's 89 frames have no earlier twin. Files: lore_catalog_runs/review_invariant_match.json. Card 197. Look-alikes page for his eye: https://claude.ai/artifact/3pv5CHWbMNH4SmcGT7Q4Ts (db collection lookalikes_1). Not yet tried: matching on eyes only, or on the figure with the background masked out. ## D85 — The collector's look-alikes review (2026-09-20) He marked none of the computed look-alikes a real match (for #450 he picked #11 as closest on shape, without the key image and with bigger eyes). His direction: match on SHAPE AND MOVEMENT of the figure, not the whole frame, and search the right drop for each so-GAN: #453 against first-drop skulls (it may be a colour skull made greyscale); #463 by the timing of its colour-flash effect; #471 and #487 against the Solstice drop (Friendly Clown #203 is nearest for #471); #494 and #502 against the Ghost-drop skulls; #499 against the robo drop; #508 against the other skull/robo hybrids. Oddities he noted: #484 is a Dancing that behaves like an early Ghost (it leaves the frame); #502 is a cronenGAN that is not huge or malformed; #508's eyes are never the robo dot. All twelve notes are in the ledger. ## D86 — Shape-and-movement matching of the so-GANs, by the drop the collector named (measurement, 2026-09-20) Figure mask = pixels more than 70 (of 255) from the frame's border colour, on the stored 32×32 frames; shape score = mean over the so-GAN's frames of the best mask match in the candidate; movement = circular correlation of the mask's area and centre over the loop. Candidates limited per his notes. Result: no identical silhouette (a copy would be ~0.95+). Best shape scores: #494 → #423 0.86 (then #360 0.82, #362, #400); #479 → #287 0.84 (the Solstice backwards-Fibonacci piece) and #299 0.83; #463 → #153 0.83; #471 → #265 0.81, #263 0.81; #487 → #287 / #295 0.78; #502 → #357 0.76, #406 0.76; #450 → #54 0.76, #45 0.76, #11 0.75 (his pick) ; #508 → #187 0.73, best shape+movement #149 Silverbot; #458 → #150 0.70; #499 → #154 0.62; #453 → #56 0.60 (weak: no first-drop skull shares its shape, so 'a colour skull made greyscale' is not supported). Scores sit only a little above each drop's median (0.45–0.74), so these are family resemblances. Colour-flash timing for #463 could not be pinned by machine on 32×32 frames; needs frame sheets for his eye (#463, #462, #464, #497, #130, #133). File: lore_catalog_runs/so_shape_match.json. Nothing published while he is mid-review. ## D87 — The collector's Solution Block Review, and what it set in motion (2026-09-20) - **Rulings.** #450 reads as a key (frames 45–47; chose 'key'). Two new deliberate pairs: **P12 {411, 450}** (key / solution) and **P13 {256, 479}** (bike / song); pairs are now thirteen, and #479 sits in two. Working hypothesis, explicitly not a rule: when a Review piece matches an earlier piece, SWAP in the earlier word, then try Act 1 order, full collection order and the common variants. He is not sold on the 'drop the leading 1' binary-counter match. sound/echo feels intentional; soon/exact read as opposites; 'sort' may be an ordering instruction; no pattern seen across the working groups. - **Answers to his questions.** Other Cronenberg monsterGAN: one, #372 'common'. Correcting the species swap in #502's title gives 'Inverted skullGAN', which matches no earlier title (nor do the nine other Review mismatches). 'south' is the only compass direction among the 512 essences ('bottom' is #287). Skull/robo hybrids in the robo drop: #161, #175, #187, #191, all four already glitch carriers. #458 and #130 are both second in their binary run (…000001). In-sequence word for slot #471: 'truth' (#474), if group 11 followed groups 09 and 10. Unusual frame counts now state both numbers in the tag. - **New matching method**: a Review piece and an earlier piece that are the only two with the same species, Type and SubType. Unique: 450→411 key, 479→256 bike, 484→301 exact, **494→404 crash (Not Found)**. Single-field unique: 502→300 brief, 487→372 common / 120 auction / 261 bitter. - **hyp79** (908 phrases, plant seen, no match): the swap rule in four match-list versions × six orders × three shapes (card 198); the ten unusual-frame Review pieces as universe + nine + final (card 199). - Instruction-like essences noted for a later look: sort #494, swap #402, flip #142, arrange #99, double #328, add #26, mirror #187, divide #449, erase #495, cancel #507, again #38, twice #393, begin #241, before #240, pair #376. ## D88 — Instruction words as a class; the out-of-sequence word count; pairs runs; chain glitches page (2026-09-20) - **Instruction words** are now a class (wordlist › Essences › Word sets, status candidate, six value tags by what the command does; 45 pieces) with a sibling **position words** class (24). Lists live in build_tags.py (INSTRUCTION_WORDS, POSITION_WORDS). Thirteen instruction words sit on data-glitched pieces; eleven if the two whose only glitch is wrong timing are set aside (that exclusion is arbitrary, so the honest count is thirteen). - **Out-of-sequence words, total: eleven.** Eight in the index column + three in the solution block (soup, sorry, and sort or soul). Card 200. - **hyp80** (1,272 phrases, plant seen, no match): both out-of-sequence elevens × 12 orders × both shapes; every eleven of the thirteen instruction-on-glitch pieces; the eleven as ten + final. - **hyp81** (39,456 phrases, launched detached, running at time of writing): (a) one piece from each of the TEN DATA PAIRS (P01, P02, P03, P07–P13; all 1,024 choices) as ten + final + any and universe + ten + any, 8 orders; (b) 'the eleven in the thirteen': the thirteen pairs have two overlaps (#175 in P09 and P11; #479 in P02 and P13), which merge them into exactly ELEVEN echo groups (nine pairs, two triples); one piece from each group, all 4,608 choices, 8 orders. Reading (b) mixes data and by-eye pairs and is flagged as breaking the no-mixing rule. - **In-depth review 2 published**: Chain Glitch Review https://claude.ai/artifact/8Ux5USDihLfo1SM6AvBeAT (db collection deep_review_2; doc ids how, p-, sets, art-, next). ## D89 — hyp81 finished; the chain page rewritten (2026-09-20) - **hyp81** (39,456 phrases, 5.05M seeds, plant seen, **no match**): all 1,024 one-of-each choices over the ten data pairs as ten + final + any and universe + ten + any; all 4,608 one-of-each choices over the eleven groups the thirteen pairs merge into (mixed sides, flagged). Card 202. - **Chain Glitch Review rewritten** after the collector said the first version had too much shorthand and he could not tell what he was deciding. Version 2 has no letter codes, chips or grids: each piece is written as sentences (what is different, which side is the mistake and why), one plain question per piece, each candidate list has a plain name and a one-sentence rule with its members shown, and secondary data sits behind 'More about this piece'. New standing rule in memory: feedback_review_pages_plain_sentences. ## D90 — The collector's Chain Glitch Review (2026-09-21) - **His calls on the 29.** ON PURPOSE (9): #17, #33, #39, #44, #48, #58, #173, #281, #509. NOT SURE (8): #62, #66, #73, #81, #175, #290, #309, #315. ACCIDENT OR HABIT (12): #4, #16, #47, #59, #88, #97, #100, #121, #123, #124, #128, #383. Rule of thumb: capitals, stray marks, spaces and swapped letters look random and do not count unless Pindar pointed at them in lore or hints. New tag family chain › Sheet vs chain › The collector's call. All 29 notes are in the ledger. - **The seven lists.** Yes: none. Maybe: blockchain mistakes in the first drop; the same without the stray mark (his one bet, but with capitals, stray marks, spaces and transpositions removed); his own drop rule ("worth trying"). No: not-a-simple-swap, letter-level typos, title-only, the August list. - **What his eye says about the two-word pieces** (the register we could never read): #17 glitches more than it ghosts and is not very nebulous; #39 has one eye (cyclops); #44 a Moirai echo, cannot separate Ancient from Primordial; #173 is a ghost and NOT a smoke; #281 does NOT glitch, an ordinary seasonal bitGAN; #509 skull eyes, robo mouth; #66 has no rainbow (no greens or blues) and IS flaming. Also: #58 has no frame glitch; #81 looks Flaming, not Silver; #290 has 7 eyes but never more than 5 at once; #73 says twin but shows three (Moirai echo); #100 and #44 are Moirai echoes. - **Checks he asked for.** #17's glitch IS shorter: its frame returns 4 times (2, 1, 2 frames between) against the usual 5–6 returns at 3, 2, 1, 1. #58 and #81 have no returning frame at all. #309 is Named on the sheet, and it is the only Named piece whose blockchain title uses the Name instead of the sheet title (#209, #300, #404, #441, #447, #449, #496, #512 all follow the title). Pindar's 'intentional errors' Discord message (1 Jan 2022) was a reply about #58; journal 003 names #39 the cyclopsGAN. - **hyp82** (384 phrases, plant seen, no match): his nine in the bookend shape; nine + each 'not sure' piece as ten + final and universe + ten; nine + every two of the eight as eleven. Card 203. ## D91 — The collector's review of the pairs (2026-09-21) SURE (5): P01 identical files, P02 held frames, P03 Sol and Luna, P04 scanlines, P07 TRUE and FALSE. NOT SURE (7): P05 flash pair (he reads a FAMILY of four: #130, #133, #141, #174 — zero-shaped eyes larger than the background zeros, white eyes anchored in place, a background that fills more of the square, the colour flash; this is exactly the existing tag 'Eyes are the zeros of the binary'), P06 painted numbers (a TRIO with #379 'AI'), P08 Primoridal, P09 'Gan', P10 'gltich' filenames, P11 Raibow, P12 solution and key ('I think so, but I could be over-reading it'). WITHDRAWN (1): P13 #479 with #256. Twelve pairs stand. pairs.json now carries collector_call / collector_note per pair; a withdrawn pair drops out of the tag and the site list, which is now generated from pairs.json (it had been hand-kept at eleven). His suggestion: pair pieces with exactly the same name. 37 exact two-piece names exist (lore_catalog_runs/same_title_groups.json); new data tag wordlist:same_name_pair (74 pieces). No runs: he is finishing his visual review first. ## D92 — 'Anchored zero eyes' (collector's name, 2026-09-21) The family of four (#130, #133, #141, #174) is named **anchored zero eyes**. The tag visual:white_zero_eyes keeps its id and is relabelled; its definition now describes the whole family. Pair P05's note and the brief use the name. ## D93 — 'Embedded alphanumeric' (collector's name, 2026-09-21) The painted marks are named **embedded alphanumeric**: '81' on #165, '4' on #207, 'AI' on #379 (the trio), with #179's digit-shaped boxes as the fourth tagged piece. The tag visual:painted_mark keeps its id and is relabelled. ## D94 — Same names on the website, in four versions, and what they show (2026-09-21) - **New site page /names/** (linked from the main nav): every name shared by two or more pieces, searchable by name, piece number, essence or drop; four views (errors corrected; corrected with the error pieces left out; exactly as printed on the sheet; exactly as printed on the blockchain); filters for exact pairs, names that cross drops, names with an Act 1 Review piece, names with an error piece. Error pieces always carry a gold border and a plain note saying what the error is and the collector's call. Data: pages/names/names.json (from lore_catalog_runs/same_names/site_data.json). Error piece = any of the 29 sheet-vs-blockchain disagreements plus #170 (Raibow both sides) and #205 ('Gan'). - **Counts.** Corrected: 91 shared names, 38 exact pairs. Error pieces left out: 85 names, 34 pairs. Eleven pieces have a title that differs between sheet and blockchain. Six shared names exist only because of an error piece: Silver Smoke skullGAN (#4 + #36), Rainbow Ghost skullGAN (#66 + #40), Twin Glitch skullGAN (#47, #59, #73, all three error pieces), Flaming Cronenberg skullGAN (#123 + #99), Entangled glitchGAN (#281 + #244), Ultra spaceGAN (#315 + #426). So five error pieces are one half of an exact name pair whose other half is clean: #36 afford, #40 agent, #99 arrange, #244 believe, #426 upper. G1itchb0t roboGAN becomes an exact pair (#172 robot, #182 virus) once #170 is left out. - **Names that cross drops: 20 of 91, and 19 of them involve Act 1 Review.** Seven exact cross-drop pairs (one earlier piece, one Review piece): Flaming cryptoGAN #144/#463, Inverted monsterGAN #239/#505, Dancing testGAN #301/#484, Rainbow testGAN #356/#483, Vivid monsterGAN #421/#485, Primordial monsterGAN #431/#470, Entangled testGAN #436/#481. Three names with one earlier piece and two Review pieces: Inverted glitchGAN #256/#479, #503; Rainbow monsterGAN #261/#472, #486; Entangled skullGAN #411/#450, #489. Nine Review pieces join a larger earlier family. The only cross-drop name without a Review piece: Nebulous skullGAN (#42, #62, #98 in the Skull drop, #364 'matrix' in the Ghost drop). Ghost-drop skulls that carry first-drop names: #335, #364, #423. - **Odd ones out** (lore_catalog_runs/same_names/odd_ones_out.json): 67 pieces are the single odd piece in their name group on some trait. Frame glitch: #14 is the only one of 21 Rainbow skullGANs that glitches (its file is #75's); #41 the only Silver Glitch skullGAN that glitches; #71 the only Rainbow Shimmer; #166 the only Rainbot; #390 the only Frozen ghostGAN; #479 the only Inverted glitchGAN. #461 'buyer' is odd five ways among the Slender roboGANs. #364 'matrix' is odd three ways. No runs: the collector is finishing his visual review. ## D95 — Same names as a tag category that opens the Same names page (2026-09-21) The collector wanted same names to be a category of tag, and a click to go to the web page. New sidebar category **Same names** (key `samename`, after Name & essence findings): one tag per shared name, 91 tags in four sub-groups by how many pieces share the name (exactly two; three or four; five or six; seven or more). Names are the corrected ones. A click does not filter the grid: it opens /names/?name=, which shows that name's pieces side by side with the error pieces marked and the four versions available. The tags are built in build_tags.py from lore_catalog_runs/same_names/site_data.json; each piece's card line names the other pieces with its name. The CSV export gained a Same names column. Note: the API caches /api/tags for five minutes, so a new tag group can take that long to appear. ## D96 — The idea swarm of 21 September, and running all of it (2026-09-21) - **The collector's new standing instruction:** run every theory that can be stated, with every variant of any open choice, without asking; then report in plain language what was tested and why it was thought valid. (Memory: feedback_hypothesis_before_sweep.) - **The swarm.** Five idea agents, one shared briefing (theories/swarm_2026-09-21/CONTEXT.md), five lenses: same-name groups; lying labels and instruction words; Act 1 Review as the answer key; puzzle design and cipher history; order. Reports: theories/swarm_2026-09-21/lens1..5. Ideas only; every claim re-checked by me before use. - **What came out that is worth keeping as FACT.** (1) Pindar's podGALLERY solution used the FIRST AND LAST square of each of SIX pods, reached by finding DUPLICATES, and the ends were 'out of sequential order'; Act 1 has six drops whose ends come in pairs, and the twelve end-piece essences pass the checksum as they stand. (2) Twelve Review files break the normal file timing (ten odd frame counts + the two held-frame files); in mint order their essences pass the checksum and end on 'final', and started at 'calm' they end '…final, give, random'. (3) In name families of three or more, exactly eleven pieces sit in a different drop from the rest of their family, the same eleven in all four name views; FALSE #473 is one, TRUE #363's family stays in one drop. (4) Eleven titles contradict their own trait columns. (5) Ten of the wordlist's twelve counting words are essences. (6) #479's thirteen negative frames mirror the eleven shared by #231, #256, #274. (7) Eyes takes exactly eleven values; no held eleven has all different. (8) Same-name pairs are mostly trait collisions (names are built from SubType + Type + species), and pair arithmetic on wordlist numbers is noise. - **The collector's one-off-names idea:** 147 pieces have a name nobody shares. Skull drop: 20, and with the Named and error pieces left out exactly eleven (#5, #7, #8, #13, #21, #34, #46, #55, #64, #89, #120). Spooky: eleven without #205. Robo: ten without the Named pieces. - **hyp83 + hyp84: 13,252 phrases, 1.7M seeds, plant seen in both, NO MATCH.** Twenty-five theories, cards 204–228. Plain-language write-up for the collector: https://claude.ai/artifact/UruNnLATNgfiwRRTKZ5jL5 (db collection swarm_tested). - Not runnable as stated (kept as lessons): the clean-twin pointer (fills six slots only); eyes that name a drop (needs the collector's count); the Eyes filter (a free test for any future eleven). ## D97 — The collector's review of the swarm theories, his three criteria, and the follow-up run (2026-09-21) - **Three criteria he set** (memory: bitgans_solution_criteria): (1) the answer must be CROSS-DROP, needing knowledge of other drops; a set inside one drop "can't be right" (he rejected the binary-counted Review twelve, #58's family, whole families in one drop, the even Glitch skulls, the skull one-off names). (2) It must be HARD, even for Pindar to back-solve; he would not reuse an earlier solution outright (first and last of each drop: "a bad idea"), though he might make it rhyme. (3) The last word must be the word 'random' or a truly random checksum-compliant word; an ordering that fixes a meaningful twelfth word is "de facto wrong". - **What he found interesting:** the out-of-drop eleven (order them by where they would have fallen at home); the mirrored robot numbers (try BIP order); the Glitch label that cannot be true (extend it across drops; "this is why I'm interested in the TRUE/FALSE GANs"); the misspelling swaps as an ordering principle ("we should exhaust this route"); the negative-frame stencil ("explore further"); each piece as a bit ("super interesting, extend it to the full collection"); the scrambled-count key. He lost the thread on the two solution-block theories (cards 217, 219) and asked how the so-block is derived and sits in Act 1 Review. On the printed-counter order he asked what came out of it and whether it overlaps glitch categories. - **New fact:** the eleven negative frames of #231, #256, #274 (frames 37, 38, 43, 44, 47, 48, 50, 51, 52, 53, 54 of 55, counted from 0) ARE the backwards-Fibonacci return pattern: the six return positions (37, 43, 47, 50, 52, 54, exactly where the robo glitch frame returns) each followed by the next frame. #479 carries the same eleven mirrored to the front, plus frames 26 and 27 at the middle. The two glitch kinds are one rhythm. [Corrected 28 September 2026 (D113): measured frame by frame, each of the three Solstice files has six negative frames, 37, 43, 47, 50, 52 and 54; the frame after each is an ordinary frame. #479 has seven, 0, 2, 4, 7, 11, 17 and 26, the Solstice six counted as 54 minus the frame plus 26. The six places are still exactly the robo glitch's return places, so the two glitch kinds are still one rhythm.] - **Printed-counter order, answered:** the counter restarts inside the robo, spooky and ghost drops, so ordering by it INTERLEAVES the drops (robot 000101, spooky 0o5 and ghost 5/144 all count 5). It changes the order of 12 of the 14 elevens checked. The only pieces whose counter is neither their number nor their place in the drop are the fourteen Review pieces already tagged 'Numbered in another base' (twelve binary, two hex). In the drop-rule eleven #44 and #173 (both in the collector's nine) share the counter 44. - **Glitch labels across the collection:** 69 pieces say glitch; 43 glitch, 26 do not (20 skulls; #200; #231, #256, #274, #293; #503). Ten pieces glitch without saying so: #7, #14, #31, #43, #67, #71, #113, #147, #390, #480. - **hyp85** (13,084 phrases, 1.67M seeds, plant seen, **no match**), cards 229–235: BIP order with bookends in place or pinned; out-of-drop eleven ordered by home position; timing orders on the twelve off-timing Review files cut to eleven; the stencil over every drop and the whole collection from both ends; each piece as a bit over seven cross-drop windows with six bit rules; the ten cross-collection Glitch liars; seven misspelling slips on 26 elevens × 7 word rules × 14 orders. ## D98 — Analogues of the eleven so-words in drops 1–5 (2026-09-22) Michael's idea. For each so-piece after 'solution', the closest earlier piece by data (same name; same species+Type+SubType; binary counter minus leading 1; eyes+dimensions+species) and by four machine-vision measures (full-colour frames via framevecs.npy; outline; brightness with inversion; figure mask). File: lore_catalog_runs/so_analogues.json. Notable: #463 sorry → #153 'puzzle' on ALL four visual measures (0.70/0.43/0.58/0.83) while its data matches are #144 'flame' (name) and #135 'ghost' (binary); #494 sort → #423 'cruel' on three measures (0.86 shape) vs data #404 'crash'; #458 soup → #150 'fire' on all four vs binary #130 'quantum'; #479 song → #287 'bottom' by shape (0.84) vs data #256 'bike'; #453 solve has no strong match anywhere (best 0.60). hyp86: 34,560 phrases, data side and visual side kept apart, every variant, plant seen, NO MATCH. Card 236. Review page for his eye (not yet published while he is away): lore_catalog_runs/so_analogues/ (db collection so_analogues, doc ids so-). ## D99 — The collector's skull-drop visual audit, and the naming conventions (2026-09-23) - **Audit file:** lore_catalog_runs/visual_audit_skull_2026-09-23.json (71 notes on 128 pieces); every note appended to the ledger as type visual_audit. Seven new by-eye tags built from it (fn:visual_audit:* in build_tags.py): Moirai echo (9: #11, #12, #44, #73, #78, #100, #117, #119, #123), Eye count differs from the Eyes trait (21), Shimmers not called Shimmer (#98, #102, #120, #122), Disappears like a Ghost not called Ghost (#25, #74, #115), Called Ghost does not disappear (#57, #105), Two figures merging (#34, #50), Silver or Smoke label questioned (#49, #81, #89). Larger eyes gained #18, #23, #27 (now 17). - **Checks.** His seven "says glitch, does not glitch" calls (#45, #46, #51, #56, #58, #60, #102) are all non-prime positions and all have no returning frame in the file: eye and file agree, and the prime rule explains every one. His three "glitches but is not a trait" (#31 Smoke, #71 Shimmer, #113 Cronenberg) are all prime positions that the file confirms glitching: the rule fires regardless of the Type word. #1: eyes do not move; a slow whole-image colour cycle he reads as a proto colour flash. - **Eyes trait.** Disagreements on 21 of 128 skulls, 15 where he sees more than printed, 6 fewer. The Eyes field is a printed number and a candidate 'label that lies'. - **Naming conventions** written up: theories/NAMING-CONVENTIONS.md (on the data page). Formula = SubType (+ Type) + species; holds for 373 of 444 pieces in drops 1–5; the first and Ghost drops follow it (breaks are deliberate), the robo and spooky drops use hand-made names (compounds and costumes). Twelve skull-drop breaks: the three Named, #17, #21, #34, #39, #44, #55, #73, #123, #128. ## D100 — Two sets from the skull audit, run (2026-09-23) hyp87 (2,194 phrases, plant seen, no match): the nine Moirai echoes in the bookend shape (universe or 'abandon' first, final before the random twelfth; five orders forwards and backwards); the fourteen skulls where the collector counts more eyes than printed, every eleven of the fourteen, five orders. Both are single-drop for now, so they fail the cross-drop test until the other drops are audited; run under the standing instruction and recorded. ## D101 — Shareable links with tags pre-selected (2026-09-23) The site now reads tags from the address: bitganspuzzlebook.com/?tag=&tag=&q= opens the Collection with those tags selected. The address bar updates as tags are clicked, and a "Copy link to this view" button appears beside the selected-tag chips. Same-name pages already take /names/?name=. Tested live: two tags from a link gave 30 pieces. ## D102 — The Arweave layer drop by drop (2026-09-23) Write-up theories/ARWEAVE-LAYER.md (data page). New: the Solstice drop is a calendar (Sol Day, Luna Night, five rows of 18, tributes on the equinoxes and solstices Mar 20 / Jun 21 / Sep 22 / Dec 21); every season video runs forward in frame order except summer frames 39/53/64 on #240–#242 (before, begin, behave); filename season and date disagree only on #240 (Jun 18) and #297 (Sep 14); #28 filename ancient_glitch vs published Type Skull. New data tags: file:season_frame_out_of_order, file:season_date_mismatch, file:filename_type_mismatch. Nineteen Ghost files were uploaded twice on 31 Jan (a retry). ## D103 — The eleven wrong-speed pieces, run (2026-09-23) file:wrong_speed = exactly eleven, cross-drop (robo #155 #176 #180 #182 #190; Solstice #228; Review #446 #467 #468 #469 #512), holding both random (#468) and final (#512). Never run before as a set. hyp88: 193 phrases, nine orders fwd+rev incl. BIP number, mint number, printed counter, frame count; final/random pinned every way; plant seen; NO MATCH. Card 237. ## D104 — The collector's robo-drop visual audit (2026-09-24) - **Audit file:** lore_catalog_runs/visual_audit_2026-09-24.json (skull notes unchanged, 46 new robo notes on #129–#192); every robo note appended to the ledger as type visual_audit. build_tags.py now reads this file for the collector's notes, and visual_audit_tags.json gained an `_extend` block that adds his by-eye members to existing concept tags. - **Colour flash and speed-up belong to the palindrome robots.** Every one of the eight nagaGANs shows a colour flash or a colour speed-up by eye. Outside the nagas only three robos do: #130 qubitGAN, #133, #161. That is exactly eleven. The file backs the speed-ups: six of his seven speed-up pieces (all but #133) have an unusual frame count, and the four flash-only pieces (#130, #141, #161, #174) have the normal 55. Colour flash now 10 pieces, colour speed-up 22, sine-wave flash 4 (#130, #133, #141, #174: the binary turns into waves during the flash). - **The robo Eyes trait is not a count.** The robo drop is the only drop that never prints an odd eye count (values used: 0, 2, 4, 8). He sees three eyes on fourteen robos: twelve printed 2, two printed 4 (#153, #179). #139 Twinbot shows four, printed 2; #192 CalcuclaC shows four, printed 8. The three pieces printed 0 (#131, #137, #145) have eyes drawn as binary digits ('10 10', '0000'). Eye disagreements now 37 across two drops. - **Labels the art does not support.** #145 S1-b0rg (Silver, essence silver) looks flaming; #150 Firebot (Flaming, essence fire) looks silver: the two look like each other's label, while each piece's name, SubType and essence agree with one another. Smoke questioned on #133 and #165 (both Cyber cryptoGAN). Ghost questioned on #156 and #169. Both Shimmering cryptoGANs (#143, #168) lack the slow Shimmer colour. - **Hybrids and marks.** Skull/robo hybrids by eye: #175, #187, #191 (added #187, #191 to Mixes features from two drops); #191's head recalls #130's helmet. '81' on #165 re-confirmed. New borderline mark: the word 'go' may appear on #129 (ledger only, not tagged). Noun-like eyes and whiskers on #162 as well as #147. - **Overlay kinds** (new tags): white overlay coloured where the GAN touches (7 cryptoGANs); coloured overlay that follows the GAN and comes and goes (#142, #165, #168, #185). - **New tags:** Called Shimmer, no slow colour; Eyes drawn as binary digits; the two overlay tags. Extended: Eye count differs (37), Called Ghost does not disappear (4), Silver or Smoke label questioned (7). - **hyp89** (2,210 phrases, plant seen, no match): the flash-or-speed-up eleven and the still-eyes-plus-nagas ten + final, seven word rules × seven orders fwd/rev; every eleven of the fourteen three-eye robos. Single-drop, so fails the cross-drop test; run under the standing instruction. Card 238. ## D105 — Robo eye numbers, pair P14, hybrids, and the number rules of every drop (2026-09-24) - **Eyes are not an index.** The robo Eyes trait (only 0, 2, 4, 8) matches no bit, position class, trait or tag (113 tests, best result at chance level). - **Robo layout rules found:** every fifth position from 4 is a Cyber cryptoGAN (12/12; off-rhythm Cyber cryptoGANs #165, #185); the six single-1 binary positions are all special (#130, #133, #161 colour flash; #131, #137, #145 the only cyberGANs, binary-digit eyes; #137 at 8 = 'eight'); squares from 4 are Smoke (break #145); #130 and #191 are bit-opposites and he saw the same helmet on both. Tags: Robo drop every fifth, Robo drop single 1 in binary. - **Pair P14 {145, 150}** (collector-requested): each looks like the other's Silver/Flaming label. - **Tag Skull/robo hybrid:** #161, #175, #187, #191. - **Number rules drop by drop** (theories/DROP-NUMBER-RULES.md, on the data page): the skull drop's looks follow Pindar's journal number classes. Odd non-class numbers are all Flaming (26/26). Squares are Silver or Smoke (9/10). Fibonacci numbers have 5 Dimensions (9/10; #7 and #8 look swapped). Primes glitch. Five of the seven Silver or Smoke labels the collector questioned sit on squares, so those labels come from the number. The spooky, Solstice, Ghost and Review drops show no number rule. Tags: odd non-class Flaming, square Silver or Smoke, breaks a number-class look rule. - **hyp90** (1,698 phrases, plant seen, no match): the ten breaks (#7, #8, #16, #31, #41, #55, #89, #145, #165, #185), plus variants with #1 or #150; seven word rules × seven orders fwd/rev; ten + final + any, eleven + any. Card 239. - **Follow-up (collector, 2026-09-24):** #7 and #8 made pair **P15** (Dimensions look swapped). The number-rules study and its ten breaks are kept as an OPEN hypothesis (card 239 status open), to be revisited once every drop has had the collector's visual review. - **Number rules on the site (2026-09-24, collector's request):** each drop tag now shows a "Number rules for this drop" block at the top of its view, listing each rule, how often it holds, and its breaks, marked as an open hypothesis. Source of truth: drop_rules.json (copied to pages/data/ by update_all.sh step 7b). Add every new per-drop number rule there. - **Dimensions 8 symmetry (collector: "mark that too", 2026-09-24):** robo Dimensions 8 = #129, #130, #141, #161, all still-eyes + colour flash + special binary number. #174 Evil Olive has still eyes + flash but Dimensions 4: added to drop_rules.json as the break and to the tag "Breaks a number-class look rule", which now holds exactly eleven across two drops. **hyp91** (850 phrases, plant seen, no match), card 240, kept open with card 239. ## D106 — The collector's spooky-drop visual audit (2026-09-25) - **Audit file:** lore_catalog_runs/visual_audit_2026-09-25.json (earlier notes unchanged; 16 new notes #193–#208), all ledgered as visual_audit; build_tags reads this file for notes. - **Findings.** Every spooky piece wears a costume except #206 Skullbot, which looks like a robo-drop skull/robo hybrid (costume tag +#194, #205, #208; hybrid tag +#206). Six consecutive pieces #197–#202 (0o4–0o11) show Ghost-drop features under the costume (GhostGAN eyes, binary mouth, a GhostGAN peeking out), though the spooky drop came first. #207 looks like the Review drop's Invasive bitGANs. Three eye families: large ghost-costume ring eyes (#193, #196, #201, #207), small ring eyes with square pupil on an oblong head (#194, #203, #205, #208), NounGAN-style pupils (#147, #162, #193, #207). #200 is a Twin in some frames (added to Twin). Eye count differs on #194 (4 printed, 3 seen) and #203 (2 printed, 3 seen). #202 has a fourth eye detached in a corner; #203 has a blue panel over the right two thirds of the background at times. - **Number rules (drop_rules.json):** Dimensions 5 only on Fibonacci positions in the spooky drop too (#194, #198, #206), so across the first three drops #7 is the only 5-Dimension piece off a Fibonacci position. The three ghost-costume Ghostly spookyGANs sit on Fibonacci positions 0, 3, 8; the fourth (#208, position 15) does not and looks different. - **New tags:** Ghost-drop features under the costume; NounGAN-style eyes or pupils; Small ring eyes with a square pupil; Large ring eyes of the ghost costume. No eleven emerged, so no run. - **Spooky layout (2026-09-25, collector's question):** counting the drop from 1, the four Ghostly spookyGANs sit on the squares 1, 4, 9, 16 (#193, #196, #201, #208; chance 1 in 1,820); this replaces the "three on Fibonacci, #208 breaks" reading. The hidden-type letter counts from 0 (its squares are #197 Pirate and #202 Masked). Non-Fibonacci primes from 0 are #200 (the drop's only Type Glitch) and #204; both have 3 eyes. #193 F read as "position zero carries F" (as #129), still open. Same-last-digit, mirror and neighbour pairings: nothing beyond chance. drop_rules.json updated. ## D107 — The collector's Solstice-drop visual audit, and the background rule (2026-09-27) - **Audit file:** lore_catalog_runs/visual_audit_2026-09-27.json (earlier notes unchanged; 92 new notes #209–#300), all ledgered as visual_audit; build_tags reads this file for notes. - **The background is the filename's season.** By eye: blue five-point marks (snowflakes) = winter video, green four-dot = spring, yellow squares = summer, orange dots = autumn; transition videos mix their two patterns. Agrees on 82 of 86 pieces where a colour was named (the other four name no background colour). Inverted pieces show the same pattern in negative (#213 purple, #230 red, #294 blue squares). On the two file-vs-date disagreements (#240, #297) the picture follows the FILENAME, so the date is the wrong field. **Break: #238** — 'yellow dot and green square', the two colours swapped between shapes (neighbour #237 from the same video: 'green dot and yellow square'). - **Type words describe the picture here:** all 7 'huge' pieces are Monster/Cronenberg/Cloud; all 4 'fuzzy' are Cloud. Label untrue: #298 Twin (not a twin), #281 Glitch (does not glitch by eye though the file has spliced frames). #241 'a little off' (robo eyes, skull mouth) → cross-drop hybrid. #223 fades like a Ghost. #260 maybe Moirai echo. Almost-twins: #135, #214, #217, #221, #226. - **Eye = file** on the negative glitch frames: he saw inverted glitch frames on exactly #231, #256, #274, the three file:negative_frames pieces. - **Numbers.** Four season videos supply exactly eleven frames each (winter, spring, spring-to-summer, summer-to-autumn); summer 15, autumn 12, winter-to-spring 9, autumn-to-winter 6. Glitch-Type pieces lean to prime places counting from 1 (5 of 8, p≈0.026; non-primes #244 place 36, #256, #274). No other number-class rule by printed or mint number. - **New tags:** Background colours swapped (#238); Inverted piece shows its season pattern in negative; Called Twin, does not look like a Twin; Huge by eye; Fuzzy by eye; Almost a twin. Extended: Disappears like a Ghost (+#223), Moirai echo (+#260), Mixes features from two drops (+#241), Named Glitch does not glitch (+#281). drop_rules.json Solstice entry rewritten (8 rules). - **hyp92** (4,004 phrases, plant seen, no match): the four eleven-frame season videos and the eleven calendar breaks (#213, #230, #238, #240, #241, #242, #256, #281, #294, #297, #298); seven word rules × nine orders fwd/rev, including calendar date and filename frame number. Single-drop. Card 241. - **#238 (collector, 2026-09-27):** the swapped-colours note is probably a transcription error. Tag kept as a candidate (yellow dot), removed from the rules box as a break; the background rule now has no confirmed break. The four uncounted pieces (#214, #215, #216, #294) name only the shape or the negative colour and fit the rule; the tributes and bookends were never in the count. - **Solstice layout written up (collector's request, 2026-09-27):** drop_rules.json now carries a `layout` narrative per drop (shown as "How the drop is laid out" above the rules). Solstice: calendar of five rows of 18; tribute opens rows 2–5 on equinox/solstice dates; each row = one year with dates never reversing; frames dealt in triples from eight ascending videos (one pair per tribute row: #235–236, #247–248, #271–272, #295–296; two sixes #229–234, #240–245); essences BIP 128–223 in order except #259/#268; mint order = piece order. Row 2 is the odd row (equinox start, Mar–Aug only, no autumn, the 'update' video code 44, the out-of-order summer triple). Frozen = winter only; Glowing never winter. #214–216 confirmed green dots by the collector, so the background rule fits all 86 season frames. ## D108 — The collector's Ghost-drop visual audit, and the SubType colour review page (2026-09-27) - **Audit file:** lore_catalog_runs/visual_audit_2026-09-27b.json (21 new notes #301–#444, all ledgered as visual_audit). #357 already had notes from earlier; #363 is a plain ghostGAN (collector). build_tags reads this file for notes. - **Species conventions by eye:** testGAN = test-pattern background, robo eyes, 'TEST' mouth; spaceGAN = starfield, five-bar eyes, star-dot mouth, invader badge; ghostGAN = empty white square eyes on binary; monsterGAN = pod-symbol background, invader badge, many eyes. Break: #306 ghostGAN with spaceGAN eyes (tag Eyes borrowed from another species). **Ultra = purple** (11/11); purple without the label #364, #379 (tag). Colour label unmet: #381 Bloody (pink/green), #423 Rainbow (blue/teal) (tag). #407 Mint Bot unlike anything in the drop, Review-like (tag Looks like the Review drop, before it, with #207). Speed-ups by eye #325, #327, #328, #329, #332, #342 (colour_speedup extended; #329 also colour flash). #322 first tiled-skull background (ledger). - **Layout (drop_rules.json):** species dealt in powers of two: testGAN 32; spaceGAN, monsterGAN, skullGAN 16 each; ghostGAN block 64 = 62 + the two one-offs #327 iceGAN and #407. Essences not a wordlist run; mint order reverses at #312, #360, #391. Ultra 11 and Vivid 11. - **hyp93** (1,854 phrases, plant seen, no match): Ultra eleven, Vivid eleven, every eleven of the thirteen purple pieces. Card 242. - **Review page for the collector (revised at his request):** LOCAL file lore_catalog_runs/ghost_subtype_review/index.html — all 144 pieces grouped by SubType, playing the ORIGINAL GIFs from workbench/gifs (he wants the full animation, not a chosen frame), NO machine colour analysis; fits/outlier/unsure marks and per-SubType descriptions kept in the browser, with copy and save-to-file buttons. He sends the saved JSON when done. The published artifact https://claude.ai/artifact/RHdDggb34SXsytH5mrgGfm (with colour strips) is superseded, left in place. ## D109 — The logic of the Ghost drop: a 9 × 16 colour wheel (2026-09-27) - **Source:** the collector's SubType colour review (lore_catalog_runs/ghost_subtype_review/collector_marks.json; his SubType descriptions in subtype_descriptions.json; his outliers #339 #340 #371 #380 #429 #430, unsure #359 #423 #438) plus hue measured from the frames (colour_cycle.json). - **Finding:** hue advances 22.5° per piece; period exactly 16; nine cycles; mean miss 17° (random 90°). Slot colours: 1 magenta, 2 rose, 3–4 red, 5 orange, 6 yellow, 7–8 yellow-green, 9 green, 10 spring-green, 11–12 cyan, 13 azure, 14 blue, 15–16 violet. **One colour break: #366** (orange in a rose slot). Borderline #342 #357 #364 #377 #403 #407 #435. - **Species quota per slot** (over 9 cycles): 4 ghost, 2 test, 1 space, 1 monster, 1 skull — 13/16 slots exact; slot 11 holds #327 iceGAN and #407 Mint Bot in ghost places; slots 4/5 trade a test for a ghost. spaceGAN, monsterGAN, skullGAN each run the rainbow exactly once. - **Colour-word SubTypes follow slots** (Bloody 1–3, Glowing 3–4, Shy 6–9, Vivid 7–11, Frozen 9–13, Ultra 13–16); 12 pieces off their slot (#309 #313 #357 #359 #361 #371 #377 #380 #393 #403 #429 #430); picture follows the slot on all but #357. Non-colour SubTypes spread over all slots (matches the collector: Dancing/Binary/Ethereal/Waking/Crypto/Cyber have no visual rule). Mint reversals at slots 12, 12, 11 — all cyan. - **Tags:** Breaks the Ghost colour cycle (#366), Slightly off (7, candidate), Colour-word SubType off its slot (12), One-off taking a ghostGAN slot (#327, #407), Colour outlier for its SubType (his 6), possible outlier (his 3), Skull icon on the body (#397). Ghost layout and rules rewritten in drop_rules.json. **hyp94** running on the break sets (card to follow). - **hyp94** (6,328 phrases, plant seen, no match): the eleven design breaks; every eleven of the twelve off-slot colour labels; every eleven of all fifteen break pieces. Card 243. ## D110 — The collector's Act 1 Review visual audit; every drop now audited (2026-09-27) - **Audit file:** lore_catalog_runs/visual_audit_2026-09-27c.json (45 new notes #445–#512; #461 has no key in the file). All ledgered as visual_audit; build_tags reads this file. - **Review layout (drop_rules.json):** tributes, then working groups walking the earlier drops in order (00–03 skull pairs, 04–05 robo, 06 crypto, 08 spooky, 09–12 Solstice + Ghost test/monster, then skullall / highdef / aa / bs / 13 / #512). Robo count continues 64–75 in binary across #457–464 and #497–500; spooky reprises 0x20, 0x21. Groups in pairs and fours; no 07. Colour: warm first half, cool second half, no cycle (best fit miss 64°). Inverted = black background, breaks #509, #511 (white). Early testGANs #467–469 = scrambled TEST background (the wrong-speed, saved-early three). #477/#478 carry the Solstice spring-to-summer pattern; #479/#480 beige→black. #497–500 Mint Bot derivatives. #506/#508/#510 skull-or-ghost eyes on Bot/bit. #454 four faces (Moirai). #512 superposition blur. - **Tags:** new Scrambled TEST background, Solstice season pattern reprised, Mint Bot derivative, Inverted piece on a white background; extended shimmer_label_untrue (+449), shimmer_unlabelled (+451, 452), moirai_echo (+454), ghost_label_untrue (+456), binary_digit_eyes (+459, 460), overlay_follows_gan (+464), noun_pupils (+467, 473–476), color_flash/speedup (+491), borrowed_species_eyes (+506, 508, 510). - **Ghost drop:** the collector's verbatim SubType descriptions added to the Ghost layout on the site (his request). - **hyp95 (cross-drop, passes the collector's test):** the eleven pieces whose Type/SubType word promises a look the picture does not show, by eye, leaving the Glitch word aside: Ghost-not-fading #57 #105 #156 #169 #456, Shimmer-not-shimmering #143 #168 #449, Twin-not-twin #298, colour-word unmet #381 #423 — five drops. Also every eleven of the thirteen that behave like a type they are not called (#98 #102 #120 #122 #451 #452 Shimmer; #25 #74 #115 #223 Ghost; #31 #71 #113 Glitch). Result below. - **hyp95 result** (1,012 phrases, plant seen, no match). Card 244. ## D111 — First-principles brief for a new run (2026-09-27) theories/FIRST-PRINCIPLES-BRIEF-2026-09-27.md (also on the data page and at ~/Desktop/bitGAN-first-principles-brief-2026-09-27.md): the collector's charge for a swarm that starts from the logic of each drop and looks for the logic of the whole 512 (bits and bytes, breaks, echoes, drops as a map to the Review and back, visual and numeric logic), with his three tests, the standing rules, the per-drop logic and breaks, cross-drop facts, the harness and recording steps, open cards, and the data map. Next D-number for the new chat: D112. ## D112 — The first-principles swarm: 37 jobs, 99,585 phrases, no match; the adversary's verdict (2026-09-27) - **Set-up:** theories/swarm_2026-09-27/ (also on the site under /data/swarm_2026-09-27/): the collector's brief, a mechanics file, an append-only shared board (BOARD.md, 40-plus entries), a questions file (eight questions for the collector), a shared hypothesis library (hyplib.py, mirrors build_hyp95.py) and a single queue runner (run_queue.py) so jobs never fight for the cores. Seven lens agents (bits and bytes, breaks, echoes, drops as a map to the Review, the Review as a map back, visual logic, numeric logic) and an adversary. The old registry was stale (last updated late August); the swarm keeps its own coverage file tested_union.json (registry + the site's tested-set files + every plant-validated run since), now 405,298 sets. The registry itself is untouched at the collector's request. - **Collector's rulings during the run:** agents may re-run old hypotheses and check each other's work; nothing old is deleted; the 'so' block may still be part of the solution; he will review the hypotheses himself (review page: swarm_2026-09-27/review_hypotheses.html). - **Runs:** 37 jobs, 99,585 phrases, plant seen on every job, no match; 17,254 word sets new to coverage (site file tested_sets_no_hit_2026-09-27.txt). Cards 245–281, one per job, each stating selection, word rule and order as the agent stated it before the sweep, with the adversary's view where it named the card. - **Facts new tonight that survive the adversary's check:** (1) In Pindar's own index column #259 and #268 hold each other's word: one swap, making three swapped pairs in three fields across three drops with #7/#8 (Dimensions) and #145/#150 (look). The swap register stops there. (2) The wordlist run of drops 1 and 4 has exactly twelve marks: slips at #21 and #57, skips at #215, #236, #269, #272, the four robo/skull holders of the skipped words (#141, #21, #131, #189), and the #259/#268 swap; the four skipped words are exactly the four words held out of place elsewhere (ledgered on the skip points). [Corrected 28 September 2026 (D113): the four skip points are not planted marks. They are the residue of Pindar's rule that no essence repeats anywhere in the 512, so the hand-placed marks in his runs are seven, not twelve.] Dead under seven rules, nine orders, both shapes, and again with the ruler's own expected words. (3) Pindar's January table counts skulls from 1 and robo, spooky and Solstice from 0; both zero pieces #129 and #193 carry F; the robo one-hot positions are all visually marked (flash or binary-digit eyes). Every from-zero selection is dead. (4) #449 'divide' = word 512 = #512 'final'; fifteen Review essences are words numbered 512 or less; the pointer map has no cycles, so a single rhyme, not a map. (5) #407 saved 30 Jan 2022 at the head of a Review session; #485–#492 inside the Ghost export; #467–#469 in Oct 2021. (6) The fourteen odd-base Review counters read as ASCII '@', A–K, space, '!', but A–K is forced by the robo drop's size; the octal-prefix-slip reading of 0x20/0x21 needs fewer assumptions. - **Claims that weakened under check:** lens 4's eleven once-reprised sources exists only under a names-then-traits merge (trait method alone ten, name method alone seven); the Moirai eleven is a tag boundary (thirteen by the collector's August notes); the namesake twelve depends on title tokenising and on counting filenames; the byte-aligned Review groups are near-forced by groups of two and four; the spooky Fibonacci-5 echo runs one way only (5 implies Fibonacci, chance about 1 in 16) and the brief oversold it. - **The adversary's central judgement:** not one of the 37 selections reaches eleven without an agent's choice (a tag boundary, a merge rule, from-zero-or-one doubling, twelve run as every eleven, or ten plus final). Pindar's rule gives eleven with no choice. So the night tested our counting, not his rule. The essence-as-word assumption is not disproven, but the space of statable selections under it is close to exhausted; the marginal return on 17,254 new sets was zero. - **The lead's grading of the round (to the collector):** 18 of 35 hypotheses were chains of two or more steps, 17 the old shape; only two gave a different word rule to different drops, which the brief asked for. The round found facts and closed gaps but did not produce the elegant multi-step solution the brief described. - **Open with the collector:** the eight questions in QUESTIONS.md (#43/#45 as Moirai echoes; #461's source; #129's F; which side of a mint inversion is wrong; the spooky reprise counters; 'divide' as a pointer; which pieces wear another drop's look; namesake essences as pointers); the reading rule for card 150 (how a picture becomes a number); and whether to run a second round under a stricter rule (every hypothesis a chain of two or more moves using at least two drops' own logics, statable in three sentences Pindar could have written). Next D-number: D113. ## D113 — The second swarm attempt, rounds 1 to 5: Pindar's puzzle grammar, 112 jobs, 124,940 phrases, no match (2026-09-28) - **Design:** theories/swarm_2026-09-27/ again (on the site under /data/swarm_2026-09-27/). The lead first wrote PUZZLE_GRAMMAR.md from Pindar's five solved puzzles (roboGAN, Fibonacci gallery, podGALLERY, Winter Solstice, byteGAN). Every solution has four stages: find the markers, see the pattern behind them, read the words off one channel, apply one twist and open the wallet in mint order. Every hypothesis had to fill those four stages, say what it predicts that can be checked without a wallet, and say where a choice of ours enters. CONTEXT2.md set the rounds: six tracks (markers, counts, twists, channel, unlock, visual), each with an append-only log in tracks/; an adversary who reads every result between rounds, writes rounds/round_N_review.md and may queue altered jobs of its own; the tracks change owners every two rounds (after rounds 2 and 4); ten rounds planned. From round 4 the channel and the twist had to come from Pindar's own record, not from the default rules. - **Rounds covered here:** 1 to 5. Round 1: 27 jobs (24 track, 3 adversary), 51,156 phrases. Round 2: 26 jobs (23 and 3), 31,843. Round 3: 22 jobs (21 and 1), 27,776. Round 4: 18 jobs (16 and 2), 694. Round 5: 18 jobs (17 and 1), 13,471; a nineteenth, unlock_r5_4, was emptied because every phrase in it was already in twists_r5_1. In all 112 jobs, 111 of them run, 124,940 phrases, plant seen on every job, no match. With the 27 September lenses the swarm has run 149 jobs and 224,985 phrases. - **Cards and runs:** cards 282 to 392, one per job that ran (unlock_r5_4 is described on card 382, twists_r5_1). Each card states the selection and the reason, the word rule and order as the track wrote them before the sweep, the phrase count and re-runs, and the adversary's one-line view from that round's review. 111 run rows (label "swarm0927 "). Coverage: 22,507 eleven-word sets new to the site's coverage files, written with the older sets from the swarm coverage file to pages/data/tested_sets_no_hit_2026-09-28.txt (112,786 lines). - **Facts that survived the round-1 review:** the Fibonacci-numbered words 1 to 233 are each held by exactly one piece (#1 #2 #3 #5 #8 #13 #34 #55 #89 #223 #378) and words 21, 377, 610, 987 and 1597 by none; exactly eleven Review pieces are on a prime number (#449 #457 #461 #463 #467 #479 #487 #491 #499 #503 #509); Pindar's January 2022 table corrects #32 to 'e' and #119 to 'd' and leaves #161, #187 and #193 as typed; 126 of the 128 skull letters fit the count from one, the misfits being #32 and #119; outside the skull and robo drops the sheet's Type Glitch pieces are the ten eye-glitchers plus #200 and #281. - **Facts that survived the round-2 review:** the Ghost 'ex' column (twelve 'ex' essences on every twelfth place, slipping at #397); the eight F robos sit on exactly the eight six-bit palindromes; no essence repeats anywhere in the 512, and the four words the Solstice run skips are exactly the four words in its range that earlier pieces already held; the so-block needs three removals to leave a rising run (soup, sorry, and sort or soul). Tooling: hyplib silently drops any list that is not eleven wordlist words, and the coverage file was not rebuilt after round 1, so later "never swept" counts ran low. - **Facts that survived the round-3 review:** Pindar's sheet spells 'glitch' wrong on exactly eleven pieces (#73 'Gltich' and the ten 'G1itch' robos); in the Solstice drop the sheet's SubType Inverted and the files with negative frames together give exactly eleven pieces; the six Ghost neighbour pairs, each exactly one word apart, with none two to five apart; Frozen, Vivid, Shy, Slender and Twin are wordlist words (the round-2 review had said no SubType is); 74 token numbers between 66 and 662 return Not Found; the printed binary count has exactly ten palindromes (the eight F robos, #458, #498); the robo drop's four skullGANs (#161, #175, #187, #191) all carry a typed error; on chain the skull and spooky drops carry HiddenType and no Essence, the robo drop both, the Solstice drop Essence and Date, the Ghost and Review drops Essence only. - **Facts that survived the round-4 review:** the lead's essence-run note (analysis/essence_runs_and_bit_reversal_2026-09-28.md): the robo drift (rank correlation 0.48 over 64, 0.75 over positions 0 to 55; the robo essences are hand-chosen, words 57 to 2046, not a run); the last row #185 to #192 breaks it (mean word 966 against 1,937 for row seven) and #184 to #192 were saved in their own session on 27 September 2021, 00:51 to 01:27, the night before the roboGAN puzzle; the 'so' block is the last 12 of the wordlist's 21 'so' words; the 'ex' column holds 12 of the 23 'ex' words from first to last; eleven Ghost pieces hold the word at their own number (#307 #322 #348 #368 #372 #380 #389 #404 #407 #429 #444); five of Pindar's six named or typed marks read as Fibonacci when their bits are reversed or complemented (#32, #119, #161, #187, #193; #58 misses), about 1 in 36, or 1 in 145 with reversal and complement only, found after looking. Method: Pindar's two text steps (keep only wordlist words, then read down in mint order keeping each word the first time) reproduce his published Solstice phrase word for word from the chain Date trait. Also: the Named per drop are 3, 8, 0, 7, 12, 11. - **Facts that survived the round-5 review:** Pindar's "Hidden Types" post (31 August 2021) defines the fibonacci-glitch, flags #32 and #119 and lists the eleven skulls to hold, in one text; the Review's files and sheet rows were public on Arweave on 31 January 2022, five days before the clue; the 1 January table equals the 31 January sheet on #1 to #300 except the blank skull Drop cell; exactly eleven files return their glitch frame on exactly 5, 3, 2, 1, 1 frames (#148 #152 #158 #160 #166 #170 #172 #176 #182 #190 #287), and #291 makes twelve under the detector's own full-run flag; the Tribonacci numbers give exactly eleven pieces on the /512 count; #479 (111011111) and #503 (111110111) are each other's bits reversed; no field other than the essence gives a wordlist word on all eleven pieces of any exact data-side eleven. - **The frame correction (round-5 review, re-checked by the site recorder on the GIFs):** #231, #256 and #274 each have six negative frames, 37, 43, 47, 50, 52 and 54, not eleven; the old list added the ordinary frame after each. #479 has seven, 0, 2, 4, 7, 11, 17 and 26, not thirteen; the first six are the Solstice six counted as 54 minus the frame, not 55 minus it, and the gaps between them run 1, 1, 2, 3, 5, 8, the Fibonacci forwards. The glitch rhythm grows across the drops: skull 3, 2, 1, 1; robo 5, 3, 2, 1, 1; the Solstice negatives on the same places; #479 forwards. It gives no eleven. Corrected in the negative-frames tag, the drop pages, the ledger (#231, #256, #274, #479) and beside D97's "new fact". - **The collector's ruling (28 September 2026):** he looked at #231's six negative frames (37, 43, 47, 50, 52, 54) and #479's seven (0, 2, 4, 7, 11, 17, 26) and answered, in his words: "There are no letters numbers or shapes." The faint-writing channel Pindar used in the byteGAN puzzle is closed for these frames. - **Claims that failed:** the colour eleven (the wordlist has 24 colour words, not 13); wrong speed as a visible mark (only #228 at 50 ms looks different); the podGALLERY hop (undefined on at least four of eleven on every exact set); the Review primes' chain back through the reprise map (undefined on eight of eleven); "every mark Pindar named is a mirror of a Fibonacci number" (39 of 128 skulls and 19 of 64 robos are such images, and the marks left out miss); "every robo rule glitch first shows at frame 37 of 55" (eight of fifteen); the Solstice skip points as planted marks (they are the no-repeat residue, so the ruler has seven marks, not twelve, and D112's "twelve marks" is corrected); "the collector's eye sees eleven Solstice negatives" (his words name nine, a different nine from the sheet's); the F class as eleven under every reading (the collector's D52 makes #193 a misfile, so F is eight, or ten with #458 and #498); the round-3 "7 in 100,000" odds for the first skull files (a record of Pindar's work order); "every word from t to z sits in robo rows six and seven" (four sit outside); "the robo last row holds four marked pieces" (strictly two, #187 and #191); "chance level for six neighbour pairs is two" (it is 0.19); "every 'a' word was used" ('action', word 21, is held by no piece); "the look breaks hit only #8 and #145" (#174 too); "only the Ghost drop has a rising run" (true only on a fixed stride; the robo drop's longest rising subsequence is 26 against a shuffle 99th percentile of 17); the "eleven negative frames" and "thirteen on #479"; the rhythm-eleven counter odds of "about 1 in 300" (1 in 270, or 1 in 59 with the composed operation); "the robo essences continue the wordlist run" (corrected in drop_rules.json and in the first-principles brief). - **The adversary's standing judgement after round 5:** the ownership swaps changed where markers and patterns come from but not where the words come from. Every exact eleven on record has had its essence read plainly, mirrored, hopped, filtered and glitched, and all are dead. The best marker now is markers_r5_1 (Pindar's three named sequences on the /512 count above #128, card 375), which has no channel; the best complete mechanism is unlock_r4_1 (the robo and Review Named through the roboGAN filter, card 367), whose marker is not a glitch. The next channel must come from the collector: the positive twins, the preview targets, or the eye-glitch boundary calls (#142, #211, #503, #512). - **Site changes with this entry:** five new tags under the new sub-group "Essences › Planted wordlist structures" (with the 'solution' block moved there and extended): the Ghost 'ex' column (12), Ghost essence is the word at its own number (11), Ghost neighbours holding neighbouring words (12), the robo last row that breaks the drift (8); and, as a candidate under Hidden-type letter, Pindar's typed marks whose bits read as Fibonacci reversed or complemented (5). 90 ledger entries on 84 pieces (source swarm_2026-09-28). Corrections in drop_rules.json (robo layout, skull Fibonacci-5 rule, spooky F rule; Ghost and Review layouts extended; Solstice and Review negative frames) and a Corrections section at the end of FIRST-PRINCIPLES-BRIEF-2026-09-27.md. Next D-number: D114. ## D114 — The second swarm attempt, round 6: 11 jobs, 2,767 phrases, no match; the collector closes the negative frames; the SuperRare brushstroke code (2026-09-28) - **Round 6:** ten track jobs ran: markers 3 (markers_r6_1 to _3), counts 3 (counts_r6_1 to _3), twists 1 (twists_r6_1) and unlock 3 (unlock_r6_1 to _3), 1,357 phrases in all. The adversary ran one more, adversary_r6_1, with 1,410 phrases. The planted test wallet was found on every job, and no job matched. The channel track queued nothing, because none of its three markers is an exact eleven with no choice of ours, and it declared the data-side channel lane closed. The visual track queued nothing until the collector answers. The twists track closed twists_r6_2 without a sweep (no leet title decodes to a wordlist word) and left twists_r6_3 waiting on the roboGAN puzzle files. This attempt now stands at 123 jobs and 127,707 phrases; with the 27 September lenses it is 160 jobs and 227,752 phrases. None is a hit. - **Cards and runs:** cards 393 to 403, one per job, in the same form as cards 282 to 392. 11 run rows (label "swarm0927 "). Card 403 (adversary_r6_1) carries a dated correction of its chance figure (see below). - **The collector's ruling (28 September 2026):** he looked at #231's six negative frames (37, 43, 47, 50, 52, 54) and #479's seven (0, 2, 4, 7, 11, 17, 26) and answered, in his words: "There are no letters numbers or shapes." The faint-writing channel Pindar used in the byteGAN puzzle is closed for these frames. (First recorded under D113; repeated here because round 6 was built on it.) - **The negative-frame scan, corrected on file:** lore_catalog_runs/negative_scan.json now lists six negative frames on each of #231, #256 and #274 (37, 43, 47, 50, 52, 54) and seven on #479 (0, 2, 4, 7, 11, 17, 26). The old lists (eleven frames on each Solstice file, thirteen on #479) are kept in lore_catalog_runs/negative_scan_superseded_2026-09-28.json. The counts track re-scanned all 512 GIFs in round 6 and the adversary re-ran it: only these four files carry inverted frames, and the next highest frame anywhere scores about 0.5. - **The SuperRare brushstroke code (found by the round-6 adversary, verified by the lead and re-checked by the site recorder):** Pindar's forge burned two bitGANs to make one new SuperRare bitGAN, and each forged piece's name ends in its brushstroke count. Eleven of them were minted before the Act 1 clue of 5 February 2022, numbered 0x01 to 0x0B (keys 1 to 11 of superrare_bitgans_onchain.json, Bit Mouth to Triclops); the next, 0x0C, was minted in May 2022. On all eleven, the stroke count minus 10,000 (or 20,000) is the word number of an Act 1 bitGAN's essence: in Pindar's own index column for the ten skull and robo pieces, and by the published essence for the Ghost piece #358: - Bit Mouth 11,400: 1400 'puzzle' (#153). Cyber Zombie 10,366: 366 'color' (#155). The Five Eyes 10,781: 781 'ghost' (#135). Skull Dot 20,051: 51 'alien' (#51). - The Extra Terrestrial 10,105: 105 'artist' (#105). Ink Bot 11,498: 1498 'robot' (#172). Glitch Skull 11,532: 1532 'satoshi' (#151). Robo Bobo 10,482: 482 'destroy' (#358). - The Twins 20,590: 590 'endless' (#186). The Sun Dancers 21,623: 1623 'skull' (#175). Triclops 10,069: 69 'ancient' (#69). - **Why it is not chance.** Three of the numbers (51, 69, 105) fall inside the skull run, where every word number is held by a piece, so they prove nothing. The other eight land above word 224, where only 289 of the 1,824 word numbers (15.8 percent) are held by any piece (Pindar's index column plus the published essences). Eight of eight there happens by chance about once in 2.5 million. The round-6 review's first figure, about 1 in 4 million, is corrected to this. Shifting the subtraction by one or two, or counting from 0, gives only 4 or 5 of 11. - **It stops at the clue.** The four numbered SuperRare pieces minted after the clue whose counts also fall above word 224 all miss: The Collaborator 11,464 (1464 'report'), Burn Out 10,987 (987 'kitten'), Zombie Face 11,197 (1197 'night', its own essence, held by no bitGAN) and Agentic skullGAN 1,869 ('trumpet'). None of these words is held by any piece. (LunaCat 10,101 gives 101, inside the skull run, so it tells nothing.) - **A second oddity.** Three of the pointed-at pieces were themselves burned in another pre-clue forge: #153 into Triclops, and #135 and #69 into Robo Bobo. About 0.4 would be expected by chance. - **Its plain reading is dead.** adversary_r6_1 read the eleven words (puzzle color ghost alien artist robot satoshi destroy endless skull ancient) in Pindar's mint order, the order the forge finished each pair, stroke order and word order, each both ways, with journal 009's '10,051' (alien) for The Extra Terrestrial, with the count from 0 as a control, with The Cryptographer's 10,101 ('arrive') first or last, and read the eleven pointed-at pieces through every standard rule and order: 1,410 phrases, no match (card 403). - **What is open.** Pindar never calls the SuperRare bitGANs glitches, and of the eleven pointed-at pieces only #172 glitches in the collector's workbook. Robo Bobo was minted in December 2021 and points at 'destroy', but #358, the piece that carries 'destroy', was saved on 4 January 2022, so at least one pointer names a word before any piece carried it. Whether the SuperRare bitGANs are in the Act 1 puzzle's scope is a question put to the collector (rounds/questions_for_collector_after_round_6.md, question 1). Pindar's 1 January 2022 table has a column labelled "Strokes", but it holds 0 or the Solstice dates, so it only shows a shared template; it is weak support. - **Facts that survived the round-6 review:** #503 and #512 carry no file glitch of either kind (no returning frame, no inverted frame), so those two eye-glitch calls rest on the collector's August notes, and the file backs nine of the eye-glitch eleven (#142 #211 #231 #244 #256 #274 #287 #291 #479); the forge burned exactly eleven Flaming skullGANs on the chain reading (#11 #33 #49 #66 #69 #87 #95 #111 #117 #123 #125; ten on the sheet, where #66 is Rainbow); #2 was burned on 12 October 2021 at 23:19 UTC from the vault 0x5173...7723, just before #66; the chain-side label errors standing on 1 January 2022 are exactly eleven, all skulls (#4 #16 #48 #58 #62 #66 #81 #88 #97 #100 #124); the three numeric differences in D56 all read 5 on chain; no leet title decodes to a wordlist word; only #504 of M16 prints n/512 on the sheet; the painted '81' on #165 and '4' on #207 are both Tribonacci numbers. - **Claims that failed in the round-6 review:** "the painted numbers single out the Tribonacci eleven" (81 and 4 are also both perfect squares, #4's difference is a capital letter only, and the painted-mark tag also names 'AI' on #379 and a possible '10.0' on #179); "at every later release the ten HODL skulls sat in nine wallets, none of them Pindar's" (nine wallets is right, but at the robo and spooky releases #2 sat in the vault the same track counts as Pindar's; no drop went to the holders as a group); "the eleven forge partners, read by adjacency" (wrong on two pairs: Pindar's own Burned bitGANs trait pairs #2 with #144 and #66 with the Noun GAN Study, so the true partners are ten bitGANs plus one outside item); "the forged tokens are not in the local data" (they are, in superrare_bitgans_onchain.json, and forge_mapping.md swept their eleven essences in August); the review's own first chance figure for the brushstroke code, 1 in 4 million (1 in 2.5 million, above). - **The review's verdict on rounds 7 to 10:** under the essence-and-column shape the data side is spent; every exact eleven the data gives with no choice of ours is dead. Round 7 keeps six owners for the swap but points every track at the open lines (the SuperRare code, Pindar's roboGAN method with his own cut, the eye-glitch boundary, the collector's other eye readings), with no phrase quota. Rounds 8 to 10 reshape into three tracks: the SuperRare code (one owner on the census, one on the mechanism), the collector's answers (one owner who turns each answer into a gate and a job the same day), and the checker and adversary. Stopping rule: if round 7 finds no mechanism for the code, stop sweeping the 512's own metadata and spend rounds 8 to 10 only on what the collector answers. - **Site changes with this entry:** a new candidate tag under Chain-only › Forge burns, "Pointed at by a pre-clue SuperRare brushstroke count" (11 pieces); ledger entries on the eleven pointed-at pieces and on #503 and #512 (source swarm_2026-09-28, round 6); the round-6 swarm files copied to /data/swarm_2026-09-27/; the round-6 sweeps covered 502 distinct eleven-word sets, of which 268 were new to the site's coverage files and were added to pages/data/tested_sets_no_hit_2026-09-28.txt (now 113,054 lines). Next D-number: D115. ## D115 — The second swarm attempt, rounds 7 and 8: 3 jobs, 4,307 phrases, no match; the SuperRare brushstroke code is Pindar's own solved SuperRare puzzle, not the Act 1 answer (2026-09-28) - **Round 7:** one job ran, channel_r7_1 (186 phrases, 16 of them re-runs of a set swept in August). The planted test wallet was found and there was no match. The other five tracks queued nothing, because nothing they stated passed the gates. Their work was censuses: the unlock track walked the forge and found that the code's pointers name four pieces, not eleven; the twists track re-checked the code and showed that "a small number plus a round offset" is how Pindar's own index column gives three Act 1 drops their essences, in plain sight; the counts track found that only the Strokes trait carries the code; the markers track found that the eleven pointed-at pieces share nothing that the rest of their drops lack; the visual track built gated scripts that wait for the collector's answers. After round 7 the attempt stood at 124 jobs and 127,893 phrases (161 jobs and 227,938 phrases with the 27 September lenses). - **Round 8:** two jobs ran, both dead with the plant found: channel_r8_1 (24 phrases, none re-runs) and markers_r8_1 (4,097 phrases, 25 re-runs). The twists, unlock and visual tracks queued nothing; the counts track sat round 8 out. The attempt now stands at 126 jobs and 132,014 phrases (163 jobs and 232,059 phrases with the lenses). None of them opens the Act 1 wallet. The adversary's review of round 8 (rounds/round_8_review.md) is summarised below. - **Cards and runs:** cards 404 (channel_r7_1), 405 (channel_r8_1) and 406 (markers_r8_1), in the same form as cards 282 to 403, each with the adversary's view, and three run rows (label "swarm0927 "). Card 406 carries a dated correction (#0's essence is 'universe'). - **The finding of the round-7 review: the brushstroke code is Pindar's own SuperRare puzzle, and it was solved in 2022.** Pindar's post "A Brief History of bitGANs" (20 October 2021) names the first three forged pieces with their counts and says there is "some sort of puzzle behind them". The round-7 review found how that puzzle works. It is Pindar's, it was solved in early 2022, and its phrase opens a wallet that belongs to whoever solved it. It is **not** the Act 1 answer, and the pieces its words point at are not an Act 1 marker. It is the sixth worked example of Pindar's method, after the five in PUZZLE_GRAMMAR.md (which now carries it as section 6). The method, in his four stages: - **Markers.** The twelve SuperRare bitGANs that his Act 1 Finale (7 February 2022) counts by name: The Cryptographer, made in late September 2021 on its own contract (journal 008 and the post of 2 October 2021 give it 10,101 strokes), and the eleven forged pieces 0x01 to 0x0B, minted from 19 October 2021 to 8 January 2022. The set is exact: the next forged piece, 0x0C, was minted on 23 May 2022. One member, The Cryptographer, sits outside the numbered run. - **Pattern.** Each piece carries a number printed for another reason, its brushstroke count. It looks like a record of the robot's work, but it was built: every count is 10,000 or 20,000 plus a word number. - **Channel.** The stroke count minus 10,000 (or 20,000) is a word number in the BIP-39 list, counted from 1. The Essence shown on each SuperRare piece is the red herring. - **Order.** The Cryptographer's word first, then 0x01 to 0x0B in his mint order. There is no random word: all twelve come from the series, and the last piece's number closes the checksum. - **Dates.** Pindar funded the puzzle's wallet with #198 on 26 October 2021, when only four of the twelve numbers were public, so he fixed the phrase before most of its pieces were painted. That wallet burned #198 on 17 December 2021, and #198 became half of The Sun Dancers. He deposited the prize, #175, on 8 January 2022, the day Triclops, which carries the last word, was minted. #175 left the wallet on 1 February 2022, four days before the Act 1 clue, for the wallet that also holds #107 and #130, the prizes of the roboGAN puzzle that noohp solved. So the SuperRare puzzle was most likely solved by noohp. No solution post is among Pindar's 109 texts on file. - The runner compares each phrase only with the Act 1 prize wallet, so no sweep in this attempt could have seen a hit on another wallet. - **The Act-1-only wallet rule, and why (the lead, 28 September 2026).** The round-7 review recommended checking every phrase against all 942 wallets in the transfer records. That will not be built. Testing seed phrases against wallets that belong to other people is an attack on third-party wallets, whatever the intent. The runner checks only the Act 1 prize wallet and the planted test wallet. No agent may derive an address for, look up, or test a phrase against any other wallet, including the SuperRare puzzle's wallet, and no agent may call a block explorer or chain API. The SuperRare line is closed as a wallet question. The site does not put the twelve words beside that wallet or say in the same place that they open it. What card 403 and D114 already say is unchanged; the collector decides whether to change it. - **Round-7 corrections:** - **#135 and The Sun Dancers.** A working note of the lead read the "(0110)" in The Sun Dancers' Burned bitGANs trait as roboGAN 6, which is #135. It is reneGAN 0110: Pindar's own description of the token says "reneGAN 0110 & bitGAN 0o5", and the burn wallet received reneGANs 0110 on 19 October 2021. #135 was burned only into Robo Bobo, with #69, on 14 October 2021. D114's list (#153 into Triclops, #135 and #69 into Robo Bobo) was already right and stands. - **The odds, reconciled.** The code itself: 11 of 11, and 8 of 8 above word 224, about 1 in 2.5 million when the held word numbers above 224 are Pindar's index column plus the published essences (289 of 1,824), and about 1 in 2.7 million on the column alone (287). Both describe the same finding. Only exactly 10,000 works: shifts of up to 3 either way give 3 to 6 of 11, and subtracting 9,999, 10,001 or 9,900 gives 4 of 11. Two further figures fail as separate evidence. The "band" figure (about 1 in 7 million that eleven natural counts fall in the two windows) only restates the code. The "seven of eight are robo words" figure (about 1 in 9,000) is automatic, because before the clue the only published essences above 224 were the 60 robo words. The cross-link figures for the three pointed-at pieces that were already in the forge (1 in 86, 1 in 100, 1 in 109 and 1 in 145) are withdrawn: with each pointer kept in its own drop, and a hit counted when the pointed-at piece had been burned by the clue day (30 bitGANs), about 1.05 hits are expected and three or more come up about 1 time in 14 (the channel track's 1 in 15 agrees). D114's "about 0.4 expected" used a narrower target. The cross-links are weak, and they now have a plain cause: Pindar chose words for a phrase and forged from his own holdings. - **Other round-7 claims that failed:** "the stroke field points but does not spell" (it spells); "#175 went to The Sun Dancers' co-forger" (the transfer is real, but it was the puzzle's prize deposit); and "the code stops at the clue" as a link to the Act 1 clue (it stops because its phrase was complete on 8 January 2022). - **Facts that stand:** the code names words, not pieces. #358, the piece that carries 'destroy', was saved on 4 January 2022, two weeks after Robo Bobo printed 10,482, and it is the only piece of the 512 whose word number equals its mint slot (482). Pindar set the SuperRare mint order himself: on 20 December 2021 he minted Glitch Skull, Robo Bobo and The Twins in exactly the reverse of the order their pairs were burned on 14 October, and on 8 November he minted The Extra Terrestrial before Ink Bot although Ink Bot's pair was burned first. The round-8 review corrected the lesson drawn from this: a reversal of three items comes up about 1 time in 6, and the 8 November batch is in neither burn order nor its reverse. The simpler reason is that the phrase existed before its pieces, so Pindar minted in phrase order. The lesson that survives for Act 1 is that his printed numbering is the phrase order, even where it disagrees with the order in which things happened; it is not a separate clue. - **Pindar's puzzle-wallet habit (unlock track, round 8, from local transfer records only).** Four puzzle wallets are on file: roboGAN, Solstice, SuperRare and Act 1. In each, the last deposit came from Pindar's minting wallet by plain transfer on the day the last piece of information went public: #130 on the roboGAN clue day (28 September 2021), #210 and #211 on the Solstice drop and clue day (21 December 2021), #175 on the day the SuperRare puzzle's last word was minted (8 January 2022), and #445 at 11:12 UTC on the Act 1 clue day (5 February 2022). None was topped up afterwards. Each solved wallet was emptied by plain transfer to another wallet: roboGAN on 29 September 2021, Solstice on 13 January 2022 (the day before Pindar's solution post) and SuperRare on 1 February 2022. On the Act 1 clue day none of the earlier puzzle wallets held a bitGAN. The Act 1 wallet has received only #445, and nothing has left it. By this habit the Act 1 words were readable on 5 February 2022. That day's public record included the 31 January Arweave sheet and all 68 Review GIFs, so the round-7 review's hint against glitch sets that use Review pieces does not hold; only facts that exist on the Review chain records alone (released from August 2022) were not yet public. A second lesson: Pindar fixed the SuperRare phrase before most of its words were printed, so an Act 1 word can sit on a piece made after the phrase was fixed. The track also listed twenty other wallets that got a bitGAN from Pindar by plain transfer before the clue; they cluster on drop days and look like gifts. That list is a record, not a target, and under the rule above no phrase may be tested against any of them. - **The empty Strokes slot (channel track, round 8, checked by the site recorder).** Pindar's 1 January 2022 table has a "Strokes" column. It holds 0 on all 208 rows from #1 to #208, and the Solstice Date ("Day" to "Night") on all 92 rows from #209 to #300. The table stops at #300. The 31 January 2022 sheet has no Strokes or Date column (its columns are Number, Title, Name, GANg, SubType, Type, Eyes, Dimensions, Essence, Drop and Level), and on chain only the Solstice records carry a Date trait. No Ghost or Review record gives a value in that slot anywhere on file. The SuperRare tokens list their traits as Name, Strokes, GANg, SubType, Type, Eyes, Dimensions, the same run as the table's columns, so it is one template. In both of Pindar's late-2021 phrase puzzles the words sat in that slot: the stroke count on the SuperRare pieces, and the Date on the Solstice pieces. For the drops made after 1 January 2022 the slot is empty. That counts against an Act 1 carrier there. - **The three elevens Pindar names in his posts (markers track, round 8, as corrected by the round-8 review).** Up to the Act 1 Finale, three series that Pindar counts by name have eleven members: - the skulls with unique names at Fibonacci numbers, 0/128 to 89/128 (journal 003, 4 September 2021), after which he names #39 the cyclopsGAN as the one other uniquely named skull and writes "Not sure where that fits in." #0 is outside the 512. This is the same eleven he told holders to keep (card 93). It ran the SuperRare way as markers_r8_1 (card 406) and is dead. - "So 11 bitGANs have materialized in my CryptoVoxels Gallery" (27 August 2021), with an Interdimensional skullGAN at the centre of the room. The markers track wrote that its members are not on file. The round-8 review showed that Pindar's own words give them: his post of 31 August 2021 says the gallery password "was the order of the NFTs in the gallery", 01123581321345589. Those are the twelve Fibonacci terms 0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55 and 89, so the gallery is the named Fibonacci set with #0 at the centre. Every reading of that set has run dead, including markers_r8_1. The line stays open only if the collector remembers different pieces on those walls. - "Another 11 just appeared to create a total of 33 new bitGANs" (24 September 2021), after "22 more bitGANs just escaped from my network and minted themselves on OpenSea". The markers track read the Discord GAN Bot as a twelfth item. The round-8 review corrected this: the post says "we have seen the last one before. It is the Discord GAN Bot", so the Bot is the last of the eleven, not a twelfth. The members are not on file, and mint-slot order cannot give them without a choice of ours (it puts Dr Awkward #141 and Taco Cat #147 inside the first 22 but Star Rats #162, Evil Olive #174 and Flow Wolf #180 outside the first 33). An eleven whose last member is "one we have seen before" is the nearest wording in his posts to "eleven glitches ending with a random word", but it is only a resemblance, and only the collector can name the members (QUESTIONS.md, unlock round 3). The three series he counts as twelve (the podGALLERY twelve, the Solstice "12 specific bitGANs" and the 12 SuperRare bitGANs) are all solved puzzles. One more correction: the markers track wrote that #0 has no essence; the 31 January 2022 sheet gives #0 'universe' (word 1903). It did no harm, because markers_r8_1 tried every word in first place. - **Other round-8 results.** No number family in Pindar's 109 texts behaves like the stroke counts: read the SuperRare way, his other numbers land at chance (channel_r8_census.py). The collector's painted and shown numbers (#165, #207, #179, #131) read the same way land on nothing above word 224, and only four pieces carry one (visual_r8_1_build.py). The twists track wrote the example up as a method (analysis/superrare_worked_example.md, with no address and no phrase beside any address) and set a checker's list of five features an Act 1 job should copy: a series Pindar counts by name, a number printed for another reason, a round offset read as a word counted from 1, his mint or numbering order, and a last word that only closes the checksum. By that track's arithmetic (twists_r8_end_check.py, re-run and confirmed by the round-8 review; not re-run by the site recorder), 'ancient' is one of 128 words that close the checksum after the first eleven, and Triclops cannot be shown to have been held back to place it. The visual track marked the SuperRare painting questions (The Sun Dancers sheet, the Skull Dot question and the two canvas checks) as no longer deciding anything. - **The round-8 review (adversary, 28 September 2026).** Both round-8 jobs are closed: channel_r8_1 fails the timing rule, because the Finale was posted two days after the prize went in, and markers_r8_1 also covers the gallery eleven. The review checked the round-8 claims against local records only and derived no address, tested no phrase and looked up no wallet. It confirmed the four-wallet timing on the local records, with two limits: there are only four cases, and the record holds bitGAN movements only, so "never topped up" means "never topped up with a bitGAN". It confirmed that no Ghost or Review chain record has a Strokes or Date key, and added that no chain trait key appears on exactly eleven records (the counts are 1, 41, 92, 208, 368, 384, 510 and 512). It confirmed the posts census, with one step made visible: before anything is removed, the window lands 11 of 34 readings against about 5.3 expected; five are stroke counts and six are drop edges (300, 444, 512), Pindar's spooky offset (800, 801) and 2062, and without the stroke counts the rest is chance. It confirmed the 128 checksum closers after the first eleven words, 28 of them Act 1 essences and 8 in the skull run, of which only #69 and #117 were burned. It found that the notes saved with both round-8 jobs claimed more of the five features than the board posts did; the board posts are right. Its strongest open lines for Act 1, best first: the offset test on the numbers printed on the 512 (total duration, per-frame delays, file size, save-time digits, mint slot), which was assigned in round 7 but never run; the collector's four visual gates; the 24 September robo eleven, which needs the collector; the 27 August gallery, which closes unless the collector remembers different pieces; and #32 and #119, still unexplained. Stack rank: markers_r8_1 enters at 3 and channel_r8_1 at 48; adversary_r6_1 stays first only because it is the one job that ran a phrase Pindar made, and it is closed as a wallet question. Round 9 is the fourth ownership swap, with six tracks, and every job must pass two gates: it copies at least three of the five features with no choice of ours, and every word it uses was public on 5 February 2022 (the 31 January sheet and GIFs, and the chain records of drops 1 to 5). Round 10 is the close. The review appended two entries to QUESTIONS.md: a withdrawal of its request to check every wallet on file, and a narrower version of the CryptoVoxels gallery question. It also told the lead that the round-7 board post and rounds/round_7_review.md still print the twelve SuperRare words beside that wallet's full address, and that those lines must come out before either file is made public. - **Site changes with this entry:** the tag "Pointed at by a pre-clue SuperRare brushstroke count" now carries a dated note that the code is Pindar's own solved SuperRare puzzle and not the Act 1 answer; ledger entries on #39, #107, #130, #135, #175, #198, #210, #211, #358 and #445 (source swarm_2026-09-28, rounds 7 and 8); the round-7 and round-8 swarm files, including rounds/round_8_review.md and analysis/superrare_worked_example.md, copied to /data/swarm_2026-09-27/, except rounds/round_7_review.md, BOARD.md and QUESTIONS.md, which put the twelve words beside the SuperRare wallet's address or say there that they open it, so they stay off the site until the collector decides (the site keeps the round-6 copies of BOARD.md and QUESTIONS.md); the round-7 and round-8 sweeps covered 4,118 distinct eleven-word sets, of which 4,092 were new to the site's coverage files and were added to pages/data/tested_sets_no_hit_2026-09-28.txt (now 117,146 sets). Next D-number: D116. ## D116 — Rounds 9 and 10 of the second swarm attempt; the close (2026-09-28) - **Rounds 9 and 10:** round 9 (fourth ownership swap; counts sat out) ran one job, 4 phrases, dead. Round 10 ran no job: every remaining job is gated on the collector's answers, and he had not answered. Each track wrote a final entry (closed / open / collector only) in tracks/*.md. Totals: second attempt 127 jobs, 132,018 phrases; with the 27 September lenses 164 jobs, 232,063 phrases; tested union 432,486 sets; plant seen on every job; no match. - **Facts that held (round 9):** the 24 September 2021 "Another 11 just appeared" eleven is #151-#161 (upload times of Pindar's own images, mint slots, save times), ending on #161, which prints skull #32's '32'; on 15 February 2022 #315, #379 and #383 each went back to Pindar and returned within an hour (three of the seven Ghost pieces whose chain disagrees with the 31 January sheet; read as corrections, so those three are read as the sheet shows them); 75 pieces were still editable at or after the clue (#449-#512 and eleven others); #512's token slot cannot be shown to exist on the clue day; no piece shows Pindar's "fibonacci-glitch" (thatbloom's name for a montage effect). - **Claims that failed (round 9):** "Pindar never uses glitch for a data error" (his 1 January 2022 reply, D39, calls intentional trait errors "glitched"); 800 and 801 in the posts census were byteGAN token ends from a December 2022 post, not the spooky offset; 2062 is a book title; the round-trip pieces are three of seven disagreeing Ghost pieces, not all of them. - **Lead's rulings:** the "glitch in the portals" puzzle (November 2021) was made by the collector and promoted by Pindar, so it is not a Pindar precedent; the five-feature gate applies only to data jobs borrowing the SuperRare method, never to visual jobs gated on the collector's answers; phrases are tested against the Act 1 prize wallet only (the round-7 proposal to test against every wallet on file was refused as an attack on third-party wallets); the SuperRare puzzle wallet's address was redacted from all swarm files (unredacted copy not kept; the address remains in the local transfer records). - **Open and never run:** the offset census of the numbers printed on the 512 (counters, frame counts, delays, durations, file sizes, save-time digits, mint slots) through Pindar's round offsets; blocked by the permission system on this machine; the collector decides whether to allow it. #32 and #119 remain unexplained. - **For the collector:** final report https://claude.ai/artifact/XddkLqn6DBb1esoFcwGuV1 and question page https://claude.ai/artifact/6Vv12evgPfxAVeCdtTygPw. Decisions pending: whether to remove the SuperRare word list from card 403 and D114; whether to allow site deploys (blocked since round 8, so D116 and rounds 9-10 are local only); whether to allow the offset census. Next D-number: D117. ## D117 — The 4 October 2026 swarm so far: a new method, Phase 0 and Phase 1, and round 1 (19 jobs, 1,524 phrases, no match) (2026-10-04) - **The brief and the new method.** A new brief was written on 4 October 2026 (`theories/SWARM-BRIEF-2026-10-04.md`, copied into the swarm folder as `BRIEF.md`) from the collector's feedback on the 27 September swarm. He had stopped after reviewing rounds 1 and 2 because later rounds "seemed like they were getting worse, not better" and were "very repetitive". The new swarm builds each hypothesis from components of four kinds: ways to select pieces, ways to get a word, ways to transform a word, and ways to order the words. One hypothesis joins components into one mechanism across drops, instead of trying one move at a time. At the collector's request, a grader (`GRADER.md`) stands in for him and grades every idea against his principles (`PRINCIPLES.md`) and the clue scorecard; he does not want to grade lists himself. A pre-run gate applies to every job: composers save jobs into `proposed/`, the grader marks each one RUN, REVISE or HOLD before anything is swept, only RUN jobs move to `queue/`, and the grader reviews the results afterwards. The swarm stops after two rounds that produce nothing new. Folder: `theories/swarm_2026-10-04/`. - **Phase 0 (four scanners and a librarian).** They scanned everything the project has done, not only the last swarms. Output: `COMPONENTS.md` (837 components), `REEVALUATION.md` (399 past ideas re-evaluated: 32 to rerun, 48 to combine, 122 kept as components, 197 retired), `COVERAGE_GAPS.md` (95 gaps: components never joined, drop themes never used as readings, clues no idea used) and `CONFLICTS.md` (96 places where sources disagree). Review page: https://claude.ai/artifact/RS3ej1LLGHWz4mjHJPVAkb. - **Correction to the tested count (4 October 2026).** `tested_union.json` holds 433,178 keys, but only 289,166 are clean eleven-word sets. The other 144,012 are damaged entries: longer strings that mix label text with words, written by jobs in the 19 September style, and they do not work as "already tested" checks. So "about 432,000 sets tested" (the brief) and "tested union 432,486 sets" (D116) overstated the count. The clean figure is about 289,000 distinct eleven-word sets. Nothing was deleted; D116 stays as written. - **Facts settled against `primary_facts.py`** (where older files disagree, `primary_facts.py` wins): - Negative frames: #231, #256 and #274 have six each (frames 37, 43, 47, 50, 52, 54); #479 has seven (0, 2, 4, 7, 11, 17, 26). The older counts of eleven or thirteen shared frames, and of two frames in the round-9 census, were wrong. - 'satoshi' (#151) is a BIP-39 word (word 1532). The only essences that are not plain list words are #191 'err0r', #363 'TRUE' and #473 'FALSE'. - The robo prime rule covers 14 pieces (#136, #140, #146, #148, #152, #158, #160, #166, #170, #172, #176, #182, #188, #190), not 15. #142 (place 13, a Fibonacci number) is one of the collector's eye-glitch exceptions. - The chain-wrong pieces are 13 (#4, #16, #48, #58, #62, #66, #73, #81, #88, #97, #100, #124, #309), as D57 says. The older 16-piece list is kept as history. - The Review counts from zero. Its printed 0x20 and 0x21 only fit a count from zero. The default zero point is #449 (reading 0x20 in octal); #445 is the variant (reading it as decimal digits). - Four past stencil jobs (visual_r1_2, visual_r1_2b, adversary_r2_1 and adversary_r2_3) ran on wrong frame lists, so their misses do not cover the true stencil. Round 1 reran the stencil on the true list (eye_and_math_r1_4, fib_all_r1_2, fib_all_r1_4). - Only seven BIP-39 words are colours with no other main meaning: black, blue, brown, green, orange, pink and yellow. Each sits on one piece. Black (#262) comes from the forced Solstice run, so Pindar chose six by hand: three robots (#164 orange, #166 pink, #189 blue) and three Ghosts (#338 yellow, #378 brown, #419 green). Colour cannot supply eleven words; in the Ghost drop it has to select, order or transform. - **Phase 1: each drop read through its own theme** (six drop readers and one relationship mapper; files `phase1/drop_*.md` and `phase1/relationships.md`, each with a script). The main findings, all checked by script: - Skulls: the chain record gives exactly eleven skulls with Dimensions 5 (#1, #2, #3, #5, #7, #13, #21, #33, #34, #55, #89), because #33 reads 3 on the sheet and 5 on chain (ruled on purpose by the collector in D90). Pindar's page 005 table maps Dimensions 5 to a cube with 32 points, the literal "32" on #32. Place 119 exists only in the Ghost drop. - Robots: the eight palindrome counters are exactly the eight nagaGANs, the eight Named robots and the eight letter-F pieces; the 14 prime-not-Fibonacci counters are exactly the 14 Type Glitch robots that follow the glitch rhythm; Pindar's January table corrects #32 and #119 but keeps #161's "32" and #187's "C". - Spookies: the prize of the collector's SpookyGAN puzzle was #193 (see the correction below); his shift started at 0 and moved letters, not word-list places; the four Ghostly titles sit on the square places counted from one (#193, #196, #201, #208; about 1 in 1,820); the Review's 0x20 and 0x21 are the last two hex digits of word numbers 0x320 and 0x321, which are 'glue' and 'goat', the words of #193 and #194. - Solstice: the seven glitchers are reached two ways (the collector's eye, and typed Glitch plus re-saved on 18 December), with #281 the one red herring; #241 is Decimal 32 on Pindar's zero-based count; #240 and #297 carry dates that disagree with the season video they were cut from, and the collector's eye sides with the video; calendar row 2 breaks three rules at once. - Ghost: the colour wheel gives a Ghost word rule for the first time (the colour-negative partner, the piece opposite on the wheel, and its twin, the half turn of 1,024 places along the word list); skull #32 lands on #332 (Ultra) and #119 on #419 (Vivid), the only two eleven-piece SubTypes (about 1 in 85, noticed after the fact); all four relisted or re-minted Ghost pieces (#310, #311, #359, #390) sit in the cyan slots (about 1 in 500); the re-mints carry index-column edge words (#359 'bright' is word 225, #390 'century' is word 300). - Review: it counts from zero; its two clean repeat rules each give exactly ten earlier pieces, which fits the shape "ten words plus 'final' plus a random word"; it has exactly eleven Named pieces, ending with #512 'final'; #479's negative frames are the Solstice six counted back from frame 54, with Fibonacci gaps. - Relationships: every 'so' word minus 1,024 is an 'ex' word, and six land on the Ghost 'ex' column (rows 3 to 8); #458 'soup' points through its counter 1000001 at #130, the only piece whose description says "1024 animations"; four Review 'so' pieces repeat planted Ghost pieces (#484 repeats #301, #494 repeats #404, #499 repeats #407, #450 repeats #411); seven Ghost 'ex' words are also SLIP-39 words whose SLIP-39 numbers cover the five planted pieces between #307 and #322 (about 1 in 42); skull #32's number is carried through three bases (#32, robot #161 at 100000, Review #465 printing 0x20). - Calibration: reading a number from one drop as a place in another drop is at chance. Over 1,098 such comparisons, only two cross-drop results came out below 1 chance in 20. 74 of the 144 Ghost pieces carry some recorded mark, so "lands on a marked Ghost piece" proves little. A key-into-place step needs its own reason, not a hit count. - **Round 1.** Six composers sent 23 ideas. The grader approved 19, sent none back and held 4 (`rounds/round_1_precheck.md`). The 19 jobs held 1,524 phrases, 1,508 of them different. Each was tried with all 128 valid twelfth words, so 195,072 full phrases were checked against the Act 1 prize wallet only. The planted test phrase was found in every job. No match. Jobs: eye_and_math_r1_1, _3 and _4; fib_all_r1_1, _2 and _4; ghost_first_r1_1, _3 and _4; number_set_r1_1, _2 and _3; review_map_r1_1 to _4; theme_per_drop_r1_1 to _3. Four were component tests (one part applied to every candidate eleven): fib_all_r1_4, number_set_r1_1, ghost_first_r1_4 and eye_and_math_r1_4. Review: `rounds/round_1_review.md`. - Held, so nothing is ruled out: eye_and_math_r1_2 (the collector's "firsts"; his notes call at least three more pieces "first", #302, #322 and #485, which would make thirteen or fourteen); eye_and_math_r1_5 (the speed-up pieces on Pindar's binary count; ten or eleven depending on whether #457 speeds up, which only his eye can settle, and close to a single rule); fib_all_r1_3 (Fibonacci places inside the Ghost and Review words; one rule with no drop theme, and he called its parent "too simple"); ghost_first_r1_2 (a duplicate; its four phrases ran inside review_map_r1_1). - The main result: the Ghost colour negative does not give the words when all eleven pieces are Ghost pieces. Both forms were tried on every Ghost eleven known (Ultra, Vivid, the Flaming skulls read on the Ghost count, the HODL keys, the skull squares, the fixed points, the Fibonacci places, the 'ex' column and the Ghost Named ten). This supports the collector's guess that the Ghost drop is a step that hands a rule (which pieces, what order, what twist) to the Review, and that the Review gives the words. Round 2 is built around that (`rounds/round_2_direction.md`). - Best ideas by the grader, none of which hit: fib_all_r1_1 (Pindar's three named sequences on #129 to #512, exactly eleven pieces in five drops), theme_per_drop_r1_2 (the eleven pieces whose theme label lies) and review_map_r1_1 (Review #455 and #490 point at the eleven Flaming skulls, which open the Ghost drop). Three ideas used or explained 8 of the 11 clues on the scorecard, and three more used 7. In Phase 0 no earlier idea scored more than 3, but a different agent did that grading, so the jump should be read with care. - New relationships: exactly ten Ghost words, turned 1,024 places, become Review words, and the last pair is #310 'story' with #512 'final'; the number 32 is marked in every drop's own numbering (#32, #161, #241, #332, #465); Pindar's "Another 11" robots (#151 to #161) line up with Solstice #231 to #241 in his January table (Decimal 22 to 32); the eleven Review primes laid on the robot grid hit all four even robot palindromes (about 1 in 109, noticed after the fact); no earlier-drop selection opens more than six Review pieces through the Review's repeats. - Closed: recursive Fibonacci on printed counters (it gives 8 or 9, not 11); base patterns across the 512 (no class gives eleven); base readings used as orders; the two Fibonacci word twists on the 31 known elevens; the colour negative on Ghost-only elevens. The eye-glitch eleven has now been read with about a hundred word rules; only a change of its members or a new order with a stated reason could reopen it. - The grader counted the round as productive (new components, higher scorecards, new relationships), so the count toward stopping stays at zero. - **Correction to the brief (4 October 2026; awaiting the collector's confirmation).** The brief (Part 6) says the winner of the collector's SpookyGAN puzzle received "one of the Ghost costume GANs". The spooky reader found that the prize was #193 Ghostly spookyGAN (0o0), the spooky drop's zero piece. Pindar's post "A New Portal" (25 October 2021) says whoever solves the puzzle "will be airdropped Ghostly spookyGAN (0o0)", and #193 first left Pindar's wallet on 29 October 2021, four days after the post. So the prize does not by itself point into the Ghost drop. The brief stays as written. - **Five questions for the collector** (each decides something): 1. Does #457 speed up in colour? If yes, the speed-up pieces on Pindar's binary count are eleven, and the held idea eye_and_math_r1_5 can return as one part of a larger rule. 2. Which pieces belong in your "firsts"? Do #302, #322 and #485 belong, and do #207 ("precursor") and #501 to #504 ("prelude")? If the true list is eleven, the held "firsts" idea runs; if it is thirteen or more, it stays closed. 3. Watching the Review from #481 to #512, do any pieces look like the purple Ultra ghosts or the green Vivid ghosts? If yes, those pieces may carry the Ghost step into the Review; if none do, that bridge is closed. 4. Do Ghost #386 and #403 look true to their colour label? If yes, the "lying label" set stands at eleven; if either looks false, it is twelve or thirteen and must be rebuilt. 5. Do #458 'soup' and #130 qubitGAN look related, the way a repeat and its source do? If yes, the 1,024 link between the 'ex' and 'so' blocks rests on something a viewer can see, and it becomes the main carrier in round 2. - **Site records with this entry:** 19 card entries appended to `site_sync_pending.json` (one per swept job, in plain sentences), for the one approved sync command; card numbers are assigned when it runs. Per-piece facts from Phase 1 and round 1 appended to `theories/anomaly_ledger.json` with source "swarm_2026-10-04". No tag changes in this entry. Next D-number: D118. ## D118 — Round 2 of the 4 October 2026 swarm (15 jobs, 511 phrases, no match) (2026-10-04) - **The grader's direction for round 2** (`theories/swarm_2026-10-04/rounds/round_2_direction.md`, written in the collector's voice). Round 1 showed that a Ghost-only eleven does not give the words through the Ghost colour negative, so round 2 asked for ideas in which the Ghost drop hands the Review a rule (which pieces, what order or what twist) and the Review gives the words. The bar was raised: every whole idea had to use or explain at least six of the eleven clues, read at least three drops through their own themes, consult at least two past puzzles in named steps, state each order with its reason before the run, and get exactly eleven from its rule with no trimming. The eye-glitch eleven was closed to new word rules, and one shared coverage job (fib_all) was set to apply round 1's twists to the six round 1 elevens no component test had touched. - **What ran.** Six composers sent 21 ideas. The grader approved 15, sent 1 back for one fix and held 5 (`rounds/round_2_precheck.md`). The 15 jobs held 511 phrases, 505 of them different (six phrases appear in both fib_all_r2_5 and theme_per_drop_r2_1). Each was tried with all 128 valid twelfth words, so 65,408 full phrases were checked against the Act 1 prize wallet only. The planted test phrase was found in every job. No match. Jobs: eye_and_math_r2_2 and _3; fib_all_r2_1, _2, _4 and _5; ghost_first_r2_1 to _3; number_set_r2_1; review_map_r2_1 to _4; theme_per_drop_r2_1. fib_all_r2_5 (349 phrases) was the declared component test. Of the 15, 6 were recombinations, 4 advancements, 4 net-new and 1 the coverage job. - **The six that did not run, and why** (nothing is ruled out for them): - eye_and_math_r2_1 (HOLD): the Inverted pieces, each drop inverting its own way. It reaches only two drops and uses one past puzzle, below the round 2 bar; the composer said so first. - fib_all_r2_3 (HOLD): two of its three versions are the same phrases as review_map_r2_1 and review_map_r2_2, which ran with fib_all credited. Its third version gave one piece (#494) a rule of its own to reach eleven, which the collector's rules forbid. - number_set_r2_2 (REVISE): its second selection needed a skull inversion the skull drop never shows, and two of its phrases were already in fib_all_r2_2. With those removed it would hold six phrases; it waits for that fix and a second check. - number_set_r2_3 (HOLD): #32 and its negative in every drop. The skull negative uses an inversion the skull drop never shows, and the Review negative (#496) needs a fallback word. - theme_per_drop_r2_2 (HOLD): the lying labels judged on Pindar's own 1 January 2022 table, with the Review's hex counters added. On the chain record that rule gives thirteen; switching to the table removes exactly #32 and #119. The grader called that a rescue, not a reason. - theme_per_drop_r2_3 (HOLD): it uses the same thirteen-made-eleven selection. - **Findings checked by script this round** (counting only, from `primary_facts.py`, Pindar's table and the phase 1 modules): - Exactly eleven Review pieces carry one of the seven Ghost colour labels (Bloody, Flaming, Glowing, Shy, Vivid, Frozen, Ultra, the seven SubTypes that hold a fixed arc of the Ghost colour wheel): #452, #455, #463, #468, #475, #482, #485, #490, #491, #494 and #495 (Flaming 5, Glowing 2, Vivid 2, Frozen 2). Sheet and chain agree. Two composers (ghost_first and review_map) found this independently. Counting 'Rainbow' would add 15; Ultra and Vivid alone give two (#485, #494). - #465 (0x20) and #466 (0x21) are the only pieces in the collection that print a hex counter, so the Review prints 32 and 33. The value-32 line on Pindar's number line is #32, #161, #241, #332 and #465; the value-33 line is #33, #162, #242, #333 and #466. #465's 0x20 read in octal is 16, its own place counting the Review from #449. - The three-sequence eleven (Fibonacci, Padovan and Tribonacci numbers from #129 to #512) becomes twelve if the boundary is set after Pindar's last printed term, 89, because Padovan then adds skull #114. None of the fourteen skull numbers that are Padovan or Tribonacci but not Fibonacci carries a Name, so Pindar marked only Fibonacci in the skull drop. - The lying-label set is eleven on Pindar's 1 January table and thirteen on the chain record, once the Review's hex counters (#465, #466) count as the Review's lying label. The table removes exactly #32 and #119. - The turn of 32 in the Review's 64-piece colour pass: six of the eleven colour-labelled pieces sit 32 places from a 'so' piece (#452 to #484, #455 to #487, #482 to #450, #485 to #453, #490 to #458, #495 to #463). The chance of six or more is about 0.0035 (1 in 280), and 32 is the only one of 63 turns that gives six; the next best give four. The turn was set before looking, but the 'so' landings were noticed after, so this is a lead, not proof. The 'one colour pass' rests on a machine hue measure until the collector looks. - #504's flipped essence ('weird', word 1,994, becomes word 55, 'almost') equals its place counting the Review from #449 (55). Over all 512 pieces a flipped essence equals the piece's own place only twice (#388 and #504), and #504 is the only Inverted one. - No piece's inverted twin in its own drop holds its list-flipped word or its half-turned word, in any drop. The per-drop inversion and the word-list inversion never coincide. - Counting the Review from #445 (place p faces place 67 minus p), exactly two mirror pairs inside the 'so' block hold neighbouring words: #458 'soup' with #499 'sound', and #463 'sorry' with #494 'sort'. They include all three displaced 'so' words. Noticed after the fact. - Pindar's printed counters put #457 (1000000) at robot place 0, so only four Review primes carry a robot counter (#457, #461, #463, #499), not eleven. number_set retired number_set_r1_3 on this. - Pindar's post "A New Portal" (25 October 2021) names the prize of the collector's SpookyGAN puzzle as "Ghostly spookyGAN (0o0)", which is #193. A second agent (fib_all) read the post independently and confirmed it. The same post says the puzzle had two steps, an acrostic and then the Fibonacci letter shift; the swarm has used only the shift. The collector's confirmation is still asked for. - Of the nine robot colour labels the collector disagreed with, none passes the two-sided test (layout disagrees with the label and the picture sides with the layout), so the lying-label rule stays at eleven on the robots. - **Correction to the round 1 review (4 October 2026).** `rounds/round_1_review.md` said that the Solstice members of the Inverted chain "read through their date first and then inverted" was asked for and not built. That was wrong: it ran in eye_and_math_r1_3 as the variant "Solstice by the calendar, Review by the list" (checked by script against `queue/eye_and_math_r1_3_tasks.json`). The eye_and_math composer found the error; the grader appended a dated correction to the review, and D117 and the review text stay as written. - **Grader's view of the round.** Better in kind: six approved ideas carry a Ghost answer or rule into the Review (none did in round 1), and approved ideas score 7 to 9 of the 11 clues on the grader's count (4 to 8 in round 1). Weaker: ideas crowd onto two templates (seven small jobs on the 'so' block, three on the colour-labelled selection), several selections rest on a step the swarm took rather than one Pindar marked, and jobs count the Review from #445 or #449, whichever suits them. Top three by the grader, none of which hit: number_set_r2_1, ghost_first_r2_2 and fib_all_r2_1. - **Questions for the collector** (each decides something): Do Ghost #386 and #403 look true to their colour label (decides whether the lying-label eleven stands)? Does the Review look warm from #449 to #480 and cool from #481 to #512 (decides whether the turn of 32 rests on something a viewer sees)? Do the eleven colour-labelled Review pieces look like their label's colour (decides whether that selection is visible or only printed)? - **Site sync for round 1.** The one approved sync command succeeded for round 1 after the collector logged Wrangler in again. The 19 round 1 jobs went up as cards 409 to 427. - **Site records with this entry:** 15 card entries appended to `site_sync_pending.json` (one per swept round 2 job), for the one approved sync command; card numbers are assigned when it runs. Per-piece facts from round 2 appended to `theories/anomaly_ledger.json` with source "swarm_2026-10-04". No tag changes in this entry. Next D-number: D119. ## D119 — Round 3 of the 4 October 2026 swarm (12 jobs, 117 phrases, no match), and the collector's rulings of the day (2026-10-04) - **The collector on grading lists.** Asked to mark about 1,400 items, he wrote: "I can't grade all of these, I gave you my opinion on what makes for a good solution. I need the agents to assemble hypotheses and have a grading agent stand in my place to evaluate ideas, provide feedback, and approve what gets tested. The ideas should be recombinations or advancements of things we tried as well as net new things." So the grader (`theories/swarm_2026-10-04/GRADER.md`) stands in for him. It grades every idea against his principles (`PRINCIPLES.md`) and the clue scorecard, gives each composer feedback in his voice, and marks each job RUN, REVISE or HOLD before anything is swept. Each family sends a recombination, an advancement and a net-new idea where it can. - **New principle P18: a miss closes a combination, not its parts.** His words (typed in chat during round 3): "Remember that when a strategy doesn't work in one combination, it might work as part of another so don't be hasty in discarding strategies for GAN selection, word transformation/identification or ordering prematurely." P18 (in `PRINCIPLES.md`): a swept miss rules out only that exact eleven in those exact orders. The selection, the word rule, the transform and the order each stay in the library and may be the right part of another combination. The round 3 precheck applied P18: no job was held because one of its parts had missed somewhere else. - **Correction (4 October 2026): read "closed" as "that exact combination missed".** D117, D118, `rounds/round_1_review.md`, `rounds/round_2_review.md` and the "Closed" list in `rounds/round_3_direction.md` say that some ideas, selections or rules are "closed", "ruled out" or "retired". Under P18, each of those statements must be read as "that exact combination missed". The selections, word rules, transforms and orders used in those jobs stay available for other combinations. The same reading applies to the "This rules out ..." sentences in the round 1 and round 2 card entries in `site_sync_pending.json`. D117, D118, the reviews and the card entries stay as written; the round 3 card entries use the new wording. - **His two new families.** His words: "I'm looking at your six composer families. I think one is missing: take the 5 previous drops and use them to solve the 6th. Another is that all 6 drops each have a portion of the answer and must be combined together." They became **five_solve_sixth** and **all_six**. A third new family, **answers**, acted on his answers of the day. With eye_and_math, fib_all, ghost_first, number_set, review_map and theme_per_drop, there are now nine families (one log each in `tracks/`). - **His answers (verbatim in `COLLECTOR_FEEDBACK_2026-10-04.md`).** - #457: "Yes 457 speeds up." The speed-up pieces on Pindar's binary count are therefore eleven: robots #129, #133, #147, #159, #162, #180 and #192, and Review #457, #462, #463 and #464. - His SpookyGAN puzzle: "Yes there was an acrostic with the fibonacci based caesar shift." The acrostic step is confirmed as a past-puzzle component. - The "firsts": "I don't know what the firsts are, you need to look at the data we've saved from earlier reviews." (Settled from saved data; see below.) - He asked for a review page: "Create me an artifact with live GIFs to do any visual reviews you need." The eye review page is https://claude.ai/artifact/68eEUd9QLpXrCky9JjNWTx (full original GIFs, no machine colour hints). - Typed in chat after that page: "The review did not have warm/cool halves. The color flashes wrere the saem, but Qubit, Cyber crypto, dr. awkward and evil olive are clearly a more closely related subset inside of it." So the Review (#449 to #512) has no warm and cool halves, and the turn of 32 stays only as a numeric part. Colour flash is one kind of glitch across all twelve pieces in his notes (#129, #130, #133, #141, #161, #174, #329, #462, #463, #464, #491, #497), and four robots inside it form a closer subset: #130 qubitGAN, #133 Cyber cryptoGAN, #141 Dr Awkward and #174 Evil Olive (the lead matched his names to piece numbers from the sheet titles). - Saved on the page, on whether #458 'soup' and #130 look related like a repeat and its source: "No. They do not look related." The 1,024 link between the 'ex' and 'so' blocks rests on numbers only; it stays available as a part. - Typed in chat, on whether Ghost #386 (labelled Glowing) and #403 (labelled Flaming) look true to their colour labels: "They do look true to their labels". So the lying-label set stands at eleven by his eye: #32, #119, #161, #187, #193, #240, #297, #371, #380, #429 and #430. - **The grader's direction for round 3** (`rounds/round_3_direction.md`, in his voice). Round 3 had to change the structure of the ideas, not add variants. It named six structural directions (one mechanism that takes a different number of pieces per drop for a stated reason; a chain in which one step's answer is the next step's key; glitches as instructions; the order pass; the four clues used least, 2, 5, 8 and 9; and labelled jobs on his eye marks) and three joint builds (the Fibonacci keys sent to the latest drop that can hold them; Pindar's fibonacci-glitch as frame events; record errors as instructions). The bar: a census before every new selection, at least six of the eleven clues with two of clues 2, 5, 8 and 9, at least three drops read through their own themes, two past puzzles in named steps, and every order stated with its reason. From round 3 on, the Review body is counted from #449 as zero unless a job gives its own reason. - **What ran.** The nine families sent 18 ideas: eye_and_math 2, fib_all 2, ghost_first 1, number_set 1, review_map 3, theme_per_drop 1, answers 2, five_solve_sixth 3 and all_six 3. The grader approved 12, sent 1 back for one fix and held 5 (`rounds/round_3_precheck.md`). Of the 12: 5 advancements, 3 recombinations, 3 net-new ideas and the order pass. The 12 jobs held 117 phrases; no phrase appears in two jobs. Each was tried with all 128 valid twelfth words, so 14,976 full phrases were checked against the Act 1 prize wallet only. The planted test phrase was found in every job. No match. Jobs: ghost_first_r3_1 (3), theme_per_drop_r3_1 (4), number_set_r3_1 (41, the declared component test: the order pass), fib_all_r3_1 (2), fib_all_r3_2 (4), review_map_r3_1 (3), review_map_r3_3 (4), answers_r3_1 (24), all_six_r3_2 (12), five_solve_sixth_r3_1 (8), five_solve_sixth_r3_2 (4) and five_solve_sixth_r3_3 (8). Under P18, each miss means that exact combination did not open the wallet; its selection, word rule and order stay available. - Four jobs share one selection, Pindar's Fibonacci keys sent to the latest drop that can hold them (Review #449, #450, #451, #452, #454, #457, #462, #470, #483 and #504, counted from #449 as 0, plus Ghost #389 at 89/144), each with its own word rule: the 'divide' fold (ghost_first_r3_1), plain essence with Inverted #504 turned over (fib_all_r3_1), each piece's own negative (review_map_r3_1), and each piece read under its costume (five_solve_sixth_r3_2, which shares only #389). - The grader's top three, none of which hit: five_solve_sixth_r3_3 (Pindar's eleven deliberate marks as keys and the Review as the lock; #119's place fits only the Ghost count and opens #419 'green', labelled Vivid), ghost_first_r3_1 and answers_r3_1. - **The six that did not run, and why** (under P18 nothing is decided about their parts): - eye_and_math_r3_1 (HOLD): the inverted frame numbers read as word numbers. The eleven appears only if frames are counted from 0; Pindar's own fibonacci-glitch example counted grid positions from 1, which gives eight. Five of the eleven positions are the Solstice's six frames replayed backwards by #479, so one glitch is counted twice. - eye_and_math_r3_2 (HOLD): the August six colour-flash pieces plus the four inverted pieces. It assumed the August six are their own kind; the collector said colour flash is one kind on twelve, which makes sixteen. The closer subset of four plus the four inverted pieces makes eight. - review_map_r3_2 (HOLD): his fibonacci-glitch played on the eleven unlocked words. It consults one past puzzle where the bar asks for two, and treating a phrase as a strip of film frames is the swarm's own step. - answers_r3_2 (HOLD): the "firsts". His 26 August notes give eleven and all his saved reviews give thirteen, so the eleven depends on choosing a file. Its 72 phrases are identical to the held eye_and_math_r1_2, which stays held. - all_six_r3_1 (HOLD): Pindar's gallery password dealt across the six drops. The share each drop gets uses his SpookyGAN shift steps (0, 1, 1, 2, 3, 5), which is the swarm's step, and five settings stay open (32 versions of one idea). - all_six_r3_3 (REVISE): the per-drop assembly, 1,824 phrases. It builds new elevens by picking one mark per drop from a menu, which the direction calls stitching and which repeats the shape of the IT-124 sweep. The grader's cut leaves one eleven built by one sentence (the drop-size split, with each drop's break of its own theme), to return with a second check. - Older held jobs keep their verdicts: eye_and_math_r1_2, eye_and_math_r1_5 (superseded by answers_r3_1), fib_all_r1_3, eye_and_math_r2_1, number_set_r2_2, number_set_r2_3, theme_per_drop_r2_2, theme_per_drop_r2_3 and fib_all_r2_3. - **Facts checked by script this round** (from the tracks' censuses and the grader's precheck, against `primary_facts.py` and the phase 1 modules): - Sending each skull number class to the latest drop that can hold it, only the Fibonacci class gives exactly eleven pieces with exactly one Ghost piece (#389). #389 'correct' is word 389 (a Ghost fixed point) and its colour-negative partner is #381 'dream'. Review #467 Ethereal testGAN has #389's title and differs from it only in Dimensions. - #449 'divide' (Named 'Bean') is word 512, the piece number of its own Review negative. Dividing a word's number by 512 and keeping the remainder names a piece; outside the skull run, the pieces this fold leaves unmoved are exactly the eleven Ghost fixed points. Under the fold the 'so' block names #119 to #130 and the Ghost 'ex' column names #115 to #137. - The robo drop has exactly eight Named pieces (#129, #141, #147, #159, #162, #174, #180, #192). Six of the seven speed-up robots are among them (about 2.5 in a million by chance), and the other two Named robots, #141 and #174, are in his closer colour-flash subset. So every Named robot carries an effect his eye recorded. - Every piece in his August colour-flash category sits on Pindar's binary count (#130 at 1, #133 at 4, #462 to #464 at 69 to 71, #497 at 72); of his 27 September additions, #129, #141, #161 and #174 do too, and #329 (Ghost) and #491 (Review) do not. - Exactly eleven Review pieces carry Dimensions 6 (#454, #461, #465, #474, #484, #486, #491, #495, #496, #499, #511), and #512 is the only piece with Dimensions 11 (2 to the 11th is 2,048, the length of the word list). - The six drop sizes (128, 64, 16, 92, 144, 68) written in binary contain 1, 1, 1, 4, 2 and 2 ones, eleven in all (about 5 in 100 by chance). - The skull count holds eleven square numbers, 1 to 121, if #1 counts as a square; in Pindar's own class order #1 is Fibonacci first, which makes ten. - Mint order and piece order disagree on 280 pairs of pieces, all from eight pieces minted out of place (#24, #30, #310, #311, #359, #390, #445, #446), not on six neighbouring pairs as the round 3 direction said. - The archive sheet's row order equals piece order; the one extra archive-only row is #0 'universe'. Sheet and chain disagree on 29 pieces, and no class of those disagreements holds exactly eleven. - Breaks in Pindar's planted word runs: no rule about breaks gives eleven by itself. The nearest gives twelve and becomes eleven only by choosing 'soul' or 'sort', so it was not built. - The eleven chain errors standing on 1 January 2022 show seven kinds of move. Only one kind, reversal (#48 'wobniaR', the only reversed trait in the 512), has a word rule in the collection, so joint build 3 (record errors as instructions) was not built. - Frame events: no positional rule on frame events gives eleven. Pindar's fibonacci-glitch definition, applied to the 13 frame positions that carry an inverted frame, gives eleven counting frames from 0 and eight counting from 1. - Skull #32's and #119's numbers read in each drop's own count land on seven pieces (#32, #161, #241, #332, #481; #119, #419), not eleven. Pindar's prime rule set against the Type label, drop by drop, gives no eleven. - The archive holds 19 distinct description texts, and six shared templates cover 497 pieces, so there is no per-piece text for an acrostic to read. - **The acrostic census.** Pindar's post "A New Portal (Solved?)" (25 October 2021) describes the collector's puzzle: first the "acrostic-y" step, taking the first letter of each sentence, then the Fibonacci letter shift. The answers family took the first letters of 83 ordered lists from the collection (every phase 1 selection of eleven or twelve, the known elevens, the 'ex' column and 'so' block, the Review group names, the Named pieces' names, titles, and the sentences of the Act 1 and Review posts) and read each raw, shifted forward and shifted back: 249 strings. None is a BIP-39 word, a run of BIP-39 word prefixes or a readable instruction; English words appear at about the rate shuffled letters give. five_solve_sixth's own tries on the Review grid and group heads were also negative. The riddle's text is not saved (the post shows it only as a remote screenshot, which was not downloaded). The acrostic stays a confirmed past-puzzle component; the text it would work on has not been found, and the collector could supply his riddle's text. - **The "firsts", settled from saved data.** A script searched his saved reviews (the 26 August note file, his master spreadsheet, every September visual audit, the ledger, the site tags) for "first", "precursor", "prelude" and related words. His 26 August note file gives eleven: #187, #207, #217, #231, #256, #454, #479, #501, #502, #503 and #504. His 27 September audit adds #302 ("First SpaceGAN. ...") and #322 ("First appearance of the tiled skull icon background"), which makes thirteen. Neither file covers all 512 pieces. The August "firsts" category was filed by an agent from his notes, not by him. So the eleven depends on the choice of file, and the "firsts" jobs stay held. - **Not yet written:** the grader's post-sweep review of round 3, and with it the count toward STOP. - **Site records with this entry:** 12 card entries appended to `site_sync_pending.json` (one per swept round 3 job, describing each miss as that exact combination not opening the wallet), for the one approved sync command; card numbers are assigned when it runs. 47 per-piece records appended to `theories/anomaly_ledger.json` with source "swarm_2026-10-04": the collector's eye answers recorded as his own observations, and round 3 facts checked by script. No tag changes in this entry. Next D-number: D120. ## D120 — Round 4 of the 4 October 2026 swarm (17 jobs, 19,035 phrases, no match): the matrix, seven composed jobs, the collector's SpookyGAN solution and his Ghost eye answers (2026-10-04) - **Correction to D119 (4 October 2026).** D119 said the grader's post-sweep review of round 3 was not yet written. It is now written (`theories/swarm_2026-10-04/rounds/round_3_review.md`). It says round 3 produced new things (one new eleven by one rule across two drops, a new split of eleven across all six drops, three new word rules and one new link between Pindar's marks and the collector's eye), so the count toward STOP stayed at zero. - **The plan for round 4** (`rounds/round_4_direction.md`, the grader writing in the collector's voice). Two halves. Half A was one coded pass, the matrix. Half B was composed hypotheses from the families, with five joint builds. The direction applied P18 throughout: a miss rules out one exact combination (those pieces, read by that word rule, changed by that transform, in that order), and no part is set aside because a combination using it missed. - **The matrix: what it is and why it was built.** The collector asked that parts not be thrown away when one combination misses: "Remember that when a strategy doesn't work in one combination, it might work as part of another so don't be hasty in discarding strategies for GAN selection, word transformation/identification or ordering prematurely." He had also asked of an earlier component test: "Right idea to take many parts of solutions and apply them broadly to other selections, but was this sloppily done and not comprehensive?" The matrix is the answer. One agent built a registry (`matrix/registry.py`) of every standing part and crossed the parts in code (`matrix/build_matrix.py`), only where the collection gives a crossing a reason, and only where the exact ordered phrase had not run before. - **The registry:** 67 selections loaded (33 from rounds 1 to 3 and the collector's eye, 8 single-drop selections loaded only for word rules that leave their drop, and 26 from the Phase 0 library), 28 word rules (plus 35 theme maps and the source jobs' own readings), 13 transforms and 22 orders (11 orders and the reverse of each). Every selection gives eleven pieces, or ten plus 'final', by its own rule; `python3 matrix/registry.py --check` passed. One more selection was left out by the count, not trimmed or padded: the pieces Pindar's marks open on his printed counts give eleven landings on only ten different pieces. - **Three tiers.** Tier 1 holds the crossings the direction named with the strongest reasons (each source job's own readings, plain essence, the theme maps, the 'divide' fold, the own-drop negative where an Inverted or negative-frame member is present, the collector's Fibonacci shift in key, mint and release-then-theme order, and so on). Tier 2 holds older or weaker parts, built and labelled as separate jobs (the word-at-number and counter readings, the reverses of key and mint order, release-then-mint, release-then-piece and word-list order, and similar). Tier 3 holds crossings with no reason; they were recorded and not built. - **The coverage table** (`matrix/COVERAGE.md`, and `matrix/coverage.json` for scripts) gives every cell for every selection: 765,050 cells in all. 4,723 were built in tier 1 and 19,214 in tier 2; 2,496 had already run; 17,238 give the same phrase as another cell; 603,768 are not applicable (a part is not defined for some member); 117,611 are not motivated (tier 3). - **The 117,611 tier 3 crossings, by reason:** 53,654 read a single-drop selection by a word rule that stays inside that drop ("Fails because it is just one drop."); 22,980 are reverses of an order other than key and mint; 21,623 put the Fibonacci shift in an order other than key, mint and release-then-theme (the collector's confirmed form ran as tier 2 in the other forward orders); 13,572 flip or half-turn every word on a selection whose rule has nothing to do with inversion; 5,782 use the turn of 32 away from #32 and #465 (he saw no warm and cool halves). - **Labels.** Every phrase carries a label naming its parts: `matrix_r4 | selection | conventions | word | transform | order | shape | tier`. `matrix/MANIFEST.json` lists every combination that makes each phrase, so a hit would name all of its parts. Many phrases are made by more than one combination (1,470 of the 2,902 in block 1). - **The grader's audit** (`rounds/round_4_precheck.md` and the read-only `rounds/round_4_precheck_audit.py`). All 23,937 matrix phrases built are eleven valid seed-list words, with no duplicates across the ten jobs. None repeats any of 228,252 ordered phrases already swept (this swarm's and the 27 September swarm's queues) or a round 4 composed phrase; 682 reuse a word set swept before in another order, and all carry the label "set swept before". 600 random labels rebuilt with the builder's own engine all matched. The grader also wrote its own code from the primary facts for nine selections, nine word rules, eleven transforms and six orders: those nine selections came out identical, 60 sampled labels (six from each job) were rebuilt exactly, and a full pass over every label built only from those parts gave 5,353 exact and 0 different. No tier 3 cell was built and no gap was filled by a second rule. Two stated limits: the theme maps and the source jobs' readings were checked only by the builder's engine (which `--check` ties to phrases that actually ran), and 26 library selections carry "as coded in ..." in place of a plain rule sentence (a write-up gap, P17; it does not affect the phrases). - **What the grader cut, and why.** Three selections were cut from both tiers. The robo palindromes plus the colour-flash one-hots is a two-part rule, which P16 calls a stitched set, and its flash part (#130, #133, #161) is not the closer four the collector named. The costume eleven no longer gives eleven: its census now gives fifteen. SubType Silver is a trait value picked only because eleven pieces hold it; 73 trait values hold exactly eleven and nothing points at this one. The grader also cut #479's stencil on selections that hold no negative-frame piece (#231, #256, #274 or #479), because the stencil's positions are #479's own frame numbers and nothing in those sets points at them. In all, 4,933 phrases were cut and kept on file, not run, in `proposed/matrix_r4_NN_cut_tasks.json`: 344 from the three selections (49 tier 1 and 295 tier 2, counted from the cut files; the precheck summary gives 310) and 4,589 stencil phrases. Every cut part stays in the registry. - **What ran:** 19,004 matrix phrases in ten jobs: tier 1 in matrix_r4_01 (2,902) and matrix_r4_02 (1,772), 4,674 in all; tier 2 in matrix_r4_03 to matrix_r4_10 (1,222, 1,936, 1,521, 1,870, 2,275, 2,618, 1,682 and 1,206), 14,330 in all. Each job's write-up is `proposed/matrix_r4_NN.md`. - **The composed jobs.** Eight came in; the grader approved seven and held one (`rounds/round_4_precheck.md`). - theme_per_drop_r4_1 (8 phrases, joint build 2): the eleven lying labels, each essence moved along the list by the size of its own lie in its drop's unit (class-letter steps, days, colour-wheel slots; #32 and #161 by the number typed). - all_six_r4_1 (1 phrase): Pindar's gallery password dealt across the six drops by the drop-size split, every drop counted from 0; 89 fits no drop, so the twelfth word is random. - eye_and_math_r4_1 (8 phrases, joint build 1): the collector's eleven speed-up pieces give their binary counters, and each counter opens the piece at that place in the latest drop that can hold it (Review #449, #453, #467, #479, #482, #500, #512; Ghost #364, #369, #370, #371). The grader rated it highest. - five_solve_sixth_r4_1 (2 phrases, joint build 4): each earlier drop does one job in release order (keys, six-bit address, octal place, mint order, word), and the Review from #445 is the lock (#445, #446, #447, #448, #450, #455, #460, #470, #476, #487, #512). - review_map_r4_2 (6 phrases): the odd one out of every title family that crosses a drop (#256, #261, #364, #411, #455, #461, #467, #473, #478, #491, #497), read by three word maps. - fib_all_r4_1 (4 phrases, joint build 5): Pindar's Fibonacci keys with the Review counted from #449 as 1 (#209 Sol, nine Review pieces, #389), with and without the collector's exact shift. - fib_all_r4_2 (2 phrases): the Ghost drop's 12 by 12 grid read as the collector's own riddle (first item of each row, forward by the password); it ran as a declared test although its words come from one drop. - Kinds: three recombinations (theme_per_drop_r4_1, eye_and_math_r4_1, five_solve_sixth_r4_1), three advancements (all_six_r4_1, fib_all_r4_1, review_map_r4_2) and one net-new idea (fib_all_r4_2). The seven jobs held 31 phrases. - **Held: review_map_r4_1** (the Fibonacci keys in the spooky costume, read from the Review's far end). Its spooky step cites #465 and #466 as its demonstration, but on its own count from #449 those pieces sit at places 16 and 17, and the step does the opposite (it writes each key in octal and reads the digits as a place on the #449 count). Its write-up also says only this direction works for all eleven keys, which is choosing a reading because it gives eleven ("Modifying the theory before you even start is a red flag that this is off."). Nothing in it is ruled out; the grader asked for its opposite word rule to run in round 5 on five_solve_sixth_r4_1's lock. - **Not built, by census:** joint build 3 (the drop-size split filled by each drop's own kind of glitch) found no sentence that gives the shares 1, 1, 1, 4, 2, 2 in every drop. ghost_first proposed nothing and did census work instead. - **What ran in round 4.** 17 jobs, 19,035 phrases: the seven composed jobs (31 phrases) and the ten matrix jobs (19,004). Each phrase was tried with all 128 valid twelfth words, so 2,436,480 full phrases were checked against the Act 1 prize wallet only. The planted test phrase was found in every job. No match. Under P18 each miss means those exact combinations did not open the wallet; every selection, word rule, transform and order used stays available. - **The collector's own SpookyGAN / kilobitGAN puzzle solution (supplied 4 October 2026).** He sent a picture of his solution sheet with the words: "Here is the solution to the Spooky/Kilobitgan puzzle." It is transcribed in `theories/swarm_2026-10-04/COLLECTOR_SPOOKYGAN_PUZZLE.md` (picture: `collector_spookygan_puzzle_solution.png`). He made the puzzle in October 2021 and Pindar promoted it (Pindar's post "A New Portal", 25 October 2021; prize spooky #193). - The riddle has twelve sentences. Take the first letter of each (W, D, M, P, H, W, A, G, L, S, K, H) and move the i-th letter forward in the alphabet by the i-th of 0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55 and 89, wrapping after Z. The answer is WEN R KBITGANS ("when are kilobitGANs"); blank rows on his sheet split it into words. The lead checked all twelve letters by arithmetic. - The twelve shift numbers are exactly Pindar's Fibonacci gallery password, all twelve of them, including 89. The riddle carries its own method: "Across sticky" (acrostic), "Shift your perception" (a Caesar shift) and the path of Fibonacci are phrases in its own sentences. - The exact form at word level (the i-th word moved forward along the 2,048-word list by the i-th password number, from 0, wrapping) had already been swept in four earlier jobs: fib_all_r1_4 (152 phrases), fib_all_r2_5 (82), answers_r3_1 (8) and review_map_r3_3 (4, as "own step"). The letter form was censused in round 3 on 83 lists and found nothing. In round 4 the form ran in fib_all_r4_1, in fib_all_r4_2, and in the matrix (tier 1 in key, mint and release-then-theme order; tier 2 in the other forward orders). - **His eye on the twelve Named Ghost pieces** (saved on the eye review page, record answers/named_ghost; ticks only, no note). - Speeds up: #327 Dry iceGAN, #370 Eye Swarm, #373 Twinvader, #374 Test Test, #383 Space-clops, #401 Fire Squid, #404 404 Not Found, #407 Mint Bot, #420 Four twentyGAN and #441 Phantom. That is ten of twelve. - Flashes colour: #407 Mint Bot only (it also speeds up). He did not tick "the close kind" for it, and no piece was ticked as the close colour-flash kind. - Neither: #309 Frosty Ghost and #379 Spin Twins. - #379 and #401 carry their Names only on the chain. Among the ten Named in the 31 January archive, nine speed up and #309 does not. - **His eye on twelve unnamed Ghost pieces drawn at random** (saved on the eye review page, record answers/unnamed_ghost_sample; ticks only, no note). A script drew them from the 132 Ghost pieces with no Name, with a fixed seed (20261004): #315, #335, #340, #347, #351, #356, #357, #376, #385, #400, #411 and #433. - Speeds up: #335, #340, #347, #356, #357, #376 and #385. That is seven. - Flashes colour: #347 only (it also speeds up). - Neither: #315, #400, #411 and #433. #351: no box ticked. - **The reading.** About seven of twelve ordinary Ghost pieces speed up, against ten of twelve Named Ghost pieces. If speed-up were spread evenly over these 24 pieces, ten or more of the twelve Named ones would speed up about one time in five. So speed-up is common in the Ghost drop, as he had said (he stopped labelling it there "because so many Ghost pieces have it"), and it is at most a weak mark on the Named Ghost pieces. In the robo drop it is different: there the eight Named pieces split exactly into his two groups (six speed up; #141 and #174 are in his closer colour-flash group). Under P18 the ten Named Ghost pieces that speed up stay available as a part. - **Round 4 facts checked by script** (from the tracks' censuses and the grader's precheck): - **The octal step shown by #465 and #466, found by two families independently.** #465 is the spooky count continued to 16, which is "20" in octal; Pindar printed "0x20" on it, and it sits at place 20 of the Review counted from #445. #466 is 17, "21" in octal, printed "0x21", at place 21. five_solve_sixth and review_map each reached this step, from the same two pieces, without sharing a word set. The grader's caution: for numbers 16 to 23 the octal digits are exactly 4 more than the value, and 4 is the gap between #445 and #449, so these two pieces cannot tell the octal rule from a plain count from #449; and Pindar printed "0x", the hex prefix, not the "0o" he printed on the spooky drop. - **The odd-one-out eleven is stable across three data sources.** In every title family of three or more that spans two drops, the one member alone in its drop: exactly eleven (#256, #261, #364, #411, #455, #461, #467, #473, #478, #491, #497), the same on Pindar's sheet titles, case-kept titles and the names on the chain, and the same in the grader's own census. Two three-drop families (Rainbow skullGAN, Primordial skullGAN) have two lone members each and are left out by the stated rule; taking every lone member gives 15, and the trait-triple reading gives 17. None of the 41 Named pieces shares its title with a piece in another drop. - **No natural marker for the 1, 1, 1, 4, 2, 2 split, found by three families.** all_six, theme_per_drop and ghost_first each counted one sentence per drop against the shares (skull, robo, spooky, Solstice, Ghost, Review). "Where the drop breaks its own theme", "where the collector recorded that drop's own kind of glitch" and "where Pindar's printed rule is broken" all fail (for example theme_per_drop's counts under the last were 2, 0, 0, 2, 0, 14), and theme_per_drop also tried 41 per-piece records and more. The precheck's one-entry-per-drop cut of all_six_r3_3 fails its second check (several drops have two or three size-one breaks), so it is stitched and was not built. review_map's census found that only a strict "reprise in the source's own base" reading picks a Review pair (#465 and #466). - **Lie sizes 1, 2, 3, 5, 8.** Measured in their own drops' units, five of the eleven lies are 1 (#119), 2 (#187), 3 (#240), 5 (#193) and 8 (#297). This was found after looking. The composer put it at about 1 in 450; fib_all's check puts it at about 1 in 28, because the three letter steps are Fibonacci by chance about 64 times in 100 and the two date lies both Fibonacci about 5 times in 100. Weak evidence; no idea was built on it beyond theme_per_drop_r4_1. - **Other census results.** The collector's closer colour-flash four (#130, #133, #141, #174) are exactly the four pieces his saved words describe with a wave; nothing about their numbers picks exactly them. Among the Review's Named pieces only #509 Dark Rainbow carries a speed-up or colour-flash note in his saved notes. His saved notes name the colour of all four Ghost lying labels (#371 "Lots of green", #380 "Purple/Magenta", #429 "Purple", #430 "This is very red and orange for an Ultra."). On Pindar's printed binary count each speed-up counter names its own piece, so only the Review's place count turns the counters into keys. The #389 to #467 copy lead gave no hand-on without a choice. - **For round 5 (the grader's notes).** Register four round 4 selections in the matrix (the Fibonacci keys with the Review counted from 1, the speed-up keys' landings, the odd ones out of the name families, and five_solve_sixth's octal lock), add a twin of the Ghost Named ten built on the collector's speed-up ticks, and replace the "as coded in ..." rule text of the 26 library selections with plain sentences. Round 4's post-sweep review is not yet written, so the count toward STOP for round 4 is not yet set (the direction says the matrix alone does not count as new unless its coverage shows a relationship nobody had seen). - **Swarm totals, rounds 1 to 4** (computed from `theories/swarm_2026-10-04/results/*.json`): round 1, 19 jobs and 1,524 phrases; round 2, 15 jobs and 511 phrases; round 3, 12 jobs and 117 phrases; round 4, 17 jobs and 19,035 phrases. In all, 63 jobs and 21,187 phrases, or 2,711,936 full phrases with the 128 valid twelfth words, all against the Act 1 prize wallet only. The planted test phrase was found in all 63 jobs. No match. - **Site records with this entry:** 17 card entries appended to `site_sync_pending.json` (one per swept round 4 job; each matrix entry says what the matrix is, what its block covers, that every phrase carries a label naming its parts, and where the coverage table is), for the one approved sync command; card numbers are assigned when it runs. Per-piece records appended to `theories/anomaly_ledger.json` with source "swarm_2026-10-04": the collector's eye ticks on the twelve Named and twelve unnamed Ghost pieces, recorded as his own observations, and round 4 facts checked by script. No tag changes in this entry. Next D-number: D121. ## D121 — Round 5 of the 4 October 2026 swarm (9 jobs, 9,991 phrases, no match): five shared assumptions tested by census, a round 4 coverage gap found and repaired, and swarm totals (2026-10-04) - **Why round 5 was different** (`theories/swarm_2026-10-04/rounds/round_5_direction.md`, the grader writing in the collector's voice). The count toward STOP stood at one after round 4. Every phrase so far had rested on the same five assumptions: (1) each word comes from one piece; (2) "11 glitches" means eleven pieces; (3) the order is one of a few simple orders; (4) the phrase is eleven derived words plus a random word, not ten plus 'final' plus a random word; (5) words are read directly on the English seed-word list. Round 5 attacked those assumptions. Every family had to run a census first and report it whatever it showed, name the assumption it broke, and count its forks (how many rules it tried before one gave eleven). The matrix got a small, reasoned addition and nothing more. - **Assumption 1, a word made from a pair (review_map; census in `tracks/review_map.md`, round 5).** The census listed every pair class a stated rule defines, from primary facts and the collector's saved notes. No class holds exactly eleven pairs. Two classes hold ten. The Ghost–Review half-turn pairs (essences 1,024 apart) hold ten, but one pair is #310 'story' with #512 'final', so the class cannot be ten plus 'final' by its own rule, and the relation is the same 1,024 on every pair, so it gives one number, not ten words. The Review's reprise pairs by mutual trait triple hold ten (no #512), but no relation the collection shows gives one word per pair: #479's frame relation (frame f becomes 54 minus f) holds on none of the ten when read on the words. Other counts: confirmed swapped pairs 3 (5 with the Ghost pairs that wait on the collector's eye); exact copies 2; half-turn pairs across the collection 54; Ghost colour partners 72; the collector's by-eye pairs in his notes 7. Nothing was built. - **Assumption 1, words taken from a text (fib_all; census in `tracks/fib_all.md`, round 5; script `proposed/fib_all_r5_census.py`).** Only saved texts were used, with no web calls: 18 of Pindar's journal pages in the site data and 109 pages in the work folder. No saved Act 1 text holds exactly eleven seed-list words, eleven sentences or eleven lines by itself (journal 002 to 007, 015 and 016, the 31 August 2021 post, the Act 1 Puzzle page and the Final Drop post were all counted, both with repeats and without). Elevens inside single sentences occur at the chance rate: 33 of 1,720 sentences across the 109 pages, about 1.9 in 100. The hand-typed Names give 12, 14, 17 or 20 pieces under four readings, never eleven; the nearest, twelve, holds #512 by the same rule, so taking it out would be trimming. Nothing was built. - **Assumption 2, eleven events, frames or bits (eye_and_math, carrying the answers family; census in `tracks/eye_and_math.md`, round 5).** The frame record was settled: #231, #256 and #274 share six negative frames (37, 43, 47, 50, 52, 54), not eleven, and #479 has seven (0, 2, 4, 7, 11, 17, 26; six of them are 54 minus the Solstice six). The older "eleven" and "thirteen" counts were the frames reached by a jump in one pass of each file. No piece holds eleven glitch events of one kind when all 50 glitch files are counted the same way (visits 3 to 7, re-insertions 2 to 6, negative frames 6 or 7). "Eleven jumps in one pass" is shared by 29 files, including Pindar's own example #73, so it marks no single piece. No eleven-by-eleven layout of yes-or-no glitches can hold a phrase: Pindar's sheet has eleven columns, but its 34 glitched fields could spell only 67 of the 2,048 words. Nothing was built, and the held job eye_and_math_r3_1 was withdrawn from further rounds. - **Assumption 3, an order the collection carries (theme_per_drop; census in `tracks/theme_per_drop.md`, round 5).** Fifteen per-piece fields were checked on 13 strong selections. Only the printed counter ranks the strongest selections with no ties. Read by the count each number belongs to (the "continued count": the Review's twelve binary titles carry the robo count on from 64 to 75, and #465 and #466 carry the spooky count on to 16 and 17), it is defined, untied and new on 16 selections; it was built as theme_per_drop_r5_1. No field ranks the lying labels or the eye-glitch eleven without ties. The collector's riddle letters, the 'so' run's group order and the reprise map gave no order (empty censuses). - **Assumption 4, the census of exact tens (number_set; `proposed/number_set_r5_census.py`, `tracks/number_set.md`).** 1,092 rules in ten families were counted the same way. 38 give exactly ten pieces without #512; they describe 27 different sets, 12 of them across drops. Where the rules were not written to find elevens, a ten turns up about as often as an eleven: trait values give 5 tens and 5 elevens, title families 2 tens and 0 elevens, number classes 7 tens and 6 elevens. So reaching ten pieces is not evidence, and allowing ten plus 'final' roughly doubles the candidate selections, so each carries about half the weight of an eleven found the same way. Five tens with a pointer that does not depend on the count were registered in the matrix (the robo palindromes carried into the Review, the Named pieces whose title's GAN word differs from their GANg, the Fibonacci skulls on the contract, the ten skulls with hidden letter c, and the Fibonacci keys into the Review body). all_six_r5_1 ran one more ten and five_solve_sixth_r5_1 built another, which is held. - **Assumption 5, the SLIP-39 hop (ghost_first; census in `tracks/ghost_first.md`, round 5).** No drop holds exactly eleven SLIP-39 essences (skull 37, robo 23, spooky 2, Solstice 15, Ghost 41, Review 21). Across the matrix's 67 selections and every word rule they use (983 defined combinations), and the round 5 additions, no selection gives eleven SLIP-39 words; the best is 9 of 11. So the hop cannot be the last step of any existing selection. Reading a number on SLIP-39 and then hopping gives back the seed word at that number, so it adds nothing. The one new eleven is the hop that closes inside the joint Ghost and Review release (Ghost #301 to #444 and the Review body #449 to #512): exactly eleven pieces, #301, #311, #331, #332, #381, #415, #434, #473, #496, #505 and #509, each with one holder of its hopped word (twelve with the full Review, because #448 'legend' hops to 'dolphin' on #401). It is weak against chance (about 7 in 100) and needed two tries. It was built as ghost_first_r5_1. - **The grader's reading of the five tests** (`rounds/round_5_precheck.md`). Assumption 1 holds as far as it was tested. Assumption 2 gave nothing to build. Assumption 3 gave one narrow order. Assumption 4 showed that a ten by itself proves nothing. Assumption 5 gave one weak eleven. A census that cleanly shows an assumption holds is useful, but under the round's rule it does not count as new. - **What ran in round 5.** Nine jobs, 9,991 phrases, each tried with all 128 valid twelfth words (1,278,848 full phrases), against the Act 1 prize wallet only. The planted test phrase was found in every job. No match. - ghost_first_r5_1 (6 phrases): the hop eleven above, each word given by the hop, in three stated orders (mint order of the pieces read, release then theme, and mint order of the holders), plain and with the collector's shift as a labelled second reading. - all_six_r5_1 (4): the drop sizes in binary with the Review counted as its 64-piece body (1, 1, 1, 4, 2, 1 ones, ten in all), each 1 read by its exponent as a place, under two counts and two word readings, then 'final'. Its landings are at chance. - theme_per_drop_r5_1 (541): the continued-count order, run as a component test on 16 selections. - theme_per_drop_r5_2 (12): the lying labels moved by their lie sizes, with the four Ghost lies measured on Pindar's colour code to the colours the collector's notes name (#371 minus 2, #380 minus 4, #429 minus 4, and #430 plus 6, plus 5 or plus 4, built three ways because his answer on #430's label is pending). - matrix_r5_01 to matrix_r5_03 (1,342, 2,590 and 1,283 phrases; 5,215 in all): the round 5 additions. Eight new selections from round 4 and five tens with a pointer, review_map's opposite rule on the octal lock, and a promoted block of every all-Ghost selection read by the Ghost drop's own rules, which the collector's words motivate: "Ghost never had a puzzle of its own, so untangling the ghost drop would seem like it could be the solution or a necessary intermediate step." The registry now loads 80 selections and its check passes. 5,223 were built; the grader cut 8 (review_map's opposite rule on eight all-Review selections other than the octal lock, which stay inside one drop), kept on file in `proposed/matrix_r5_02_cut_tasks.json`; the rule stays in the registry. - matrix_r5_04 and matrix_r5_05 (1,810 and 2,403 phrases; 4,213 in all): the repair of round 4's coverage gap (next item). - Every phrase run is labelled. The grader checked that no composed phrase repeats any of 649,458 ordered phrases already swept, and that no matrix phrase repeats a phrase already run in either swarm. Every matrix label rebuilt exactly with the registry engine, and a full pass with the grader's own code, written from the primary facts, gave 1,172 exact and 0 different. `matrix/MANIFEST.json` now maps 33,373 phrases. - Under P18 each miss means those exact combinations did not open the wallet. Every selection, word rule, transform and order used stays available as a part of other combinations. - **Held: five_solve_sixth_r5_1** (4 phrases, not run). Its rule takes every piece that exchanged a value with its partner in the drops before the Review: #7 and #8, #145 and #150, #259 and #268, and, by the collector's saved SubType notes, Ghost #339 and #340 and #429 and #430. That is ten, plus 'final'. It depends on the collector's answer to round 4's question 2 (whether those two Ghost pairs look as if their colour labels were swapped). If he says no, the selection falls to six pieces and cannot make a phrase, so there is no "no" version to run now. The grader ruled that it waits; if he says yes, it can run as built. Nothing in it is ruled out. - **The round 4 matrix coverage bug, and its repair.** A theme map is a word rule that reads each drop through its own theme, and each map was named only by the drops where it differs from the default map. A bug in round 4's code made each theme map read only the drops of the first selection that used it. On later selections with other drops the map gave no word, so those cells were marked "not applicable" (with the false reason "not defined for #g") and never built. The number_set agent found the bug in round 5. The grader confirmed it by rebuilding, since the work folder has no code history: its re-creation of the round 4 behaviour predicts exactly the 190 pairs of selection and map that round 4's coverage file marked "not defined". It then rebuilt what round 4 should have built on those pairs, under round 4's tier rules and round 4's precheck cuts, and got exactly 4,213 phrases on 18 selections (1,810 tier 1 and 2,403 tier 2). These are exactly the phrases in matrix_r5_04 and matrix_r5_05, and none was in the round 4 queue. No round 4 result was wrong: all 19,004 round 4 matrix phrases ran, and every label rebuilds exactly with the corrected code. Only coverage was missing. The 18 selections include the lying labels, the speed-up eleven, the eye-glitch eleven, Pindar's three sequences, the 32 and 33 line, both drop-size bit sets and the deliberate marks. One map, the Review read by the earlier piece it repeats, made phrases for the first time; 33 of the 39 theme maps have now made phrases. Dated corrections were appended to `rounds/round_4_review.md` and `rounds/round_4_precheck.md`. - **Correction to D120 and to the round 4 site cards (4 October 2026).** D120's statements of what the matrix covered (its count of cells built, its count of cells "not applicable", and its account of each selection read through its own theme) and the round 4 matrix cards' statement that the coverage table records cell by cell what ran and why the rest did not were too wide by 4,213 phrases until matrix_r5_04 and matrix_r5_05 ran: on the 18 selections above, round 4 had skipped some or all of the theme maps while recording those cells as not applicable. The repair has now run, with no match. Two smaller corrections to D120: there are 39 theme maps, not 35 (32 made phrases in round 4, and one more only in the repair; the figure 35 was printed by the check mode, which builds only some selections' maps); and "73 trait values hold exactly eleven" could not be reproduced (Pindar's sheet values give two, SubType Silver and SubType Ultra, and number_set's census, which also counts chain traits and the hidden letter, gives five). The judgment on SubType Silver stands: nothing points at it. - **Correction: what the roboGAN puzzle's clue was.** fib_all reports, from the site's own earlier study of the puzzle pages (`pages/data/puzzle_pages.md`, section 3.1), that the roboGAN "clue" was not prose in a post. It was an image and a data file (CSV) of the roboGANs' properties ("In red below" points at a picture), and none of the clue images is saved locally. So "take the only seed words in a text" is a fair reading of Pindar's words, but no solved puzzle has yet drawn its words from the prose of a news or journal post. - **Two facts from Pindar's Act 1 Finale post (7 February 2022), from fib_all's census.** He lists "1024 podGANs" as part of Act 1, and 1,024 is the length of the SLIP-39 list his pod puzzle used (podGAN #n holds SLIP-39 word n); this supports reading the 1,024 gap between the 'ex' and 'so' blocks as "go to the SLIP-39 list", not only as a half turn of the seed list. He also counts the Review as 64 pieces ("64 Act 1 - Review bitGANs"), which is #449 to #512; this is a second source, independent of the #465 and #504 self-checks, for counting the Review body from #449. The same post splits the Ghost drop into 112 ghostGANs and 32 testGANs, and the sheet holds exactly 32 testGANs in #301 to #444. The post is dated two days after the prize went in, so clue 9 ("Only 1 clue will ever be given") argues against it as a key. - **Other round 5 facts checked by script.** Six of the eleven hopped words sit on skullGANs (#322, #397, #435, #489, #492 and #495), about 1 in 110 against nearby words; this was noticed after the rule was fixed, so it is a reason to look, not evidence. Of the four ways to pair the collector's six "out" Ghost pieces so that both of his notes come true, only one uses neighbours (#339 and #340, #429 and #430); under the six-bit wrap, only that pairing gives ten distinct Review landings (also noticed after). Review #458 (1000001) and #498 (1001001) are the two Review counters that read the same both ways on Pindar's one binary count, beside the eight Named robots. - **The count toward STOP.** It stood at one. The round 5 direction said round 5 counts as new only if it brings a new exact eleven (or ten plus 'final') with a rule for every member and a check above chance, a new word rule whose reason survives checking, or a relationship between two independent sources. The grader's precheck said none of the composed jobs counts as new unless it hits, and none hit. The grader's final verdict on round 5 and the swarm's final report are to be written in the swarm folder, `rounds/round_5_review.md` and `FINAL_REPORT.md`; they may not exist yet when this entry was written. - **Waiting on the collector's eye: three unanswered questions** (from the round 4 review). (1) Which label should #430 carry: Bloody, Flaming or Glowing? (2) Do #339 and #340, and #429 and #430, look as if their colour labels were swapped with each other? (3) Do the pieces in the Ghost drop's first column (laid out as a 12 by 12 grid) share a look? - **Swarm totals, rounds 1 to 5** (computed from `theories/swarm_2026-10-04/results/*.json`): round 1, 19 jobs and 1,524 phrases; round 2, 15 jobs and 511; round 3, 12 jobs and 117; round 4, 17 jobs and 19,035; round 5, 9 jobs and 9,991. In all, 72 jobs and 31,178 phrases, or 3,990,784 full phrases with the 128 valid twelfth words, all against the Act 1 prize wallet only. The planted test phrase was found in all 72 jobs. No match. - **Site records with this entry:** 9 card entries appended to `site_sync_pending.json` (one per swept round 5 job; the two repair entries say plainly that they ran the phrases round 4 was reported to cover and had not run), for the one approved sync command; card numbers are assigned when it runs. Per-piece records appended to `theories/anomaly_ledger.json` with source "swarm_2026-10-04" (round 5 facts checked by script). Eight tags added to `tag_schema.py` and `bitganspuzzlebook/scripts/build_tags.py`, from the tags the round 1 to 4 recorders suggested whose members are facts checked by script or the collector's own recorded eye answers (listed in `theories/swarm_2026-10-04/rounds/round_5_site_record.md`). Next D-number: D122. ## D122 — Record corrections found by the Librarian, cycle 001 of the solving system (2026-10-05) Written by the Librarian (`system/roles/librarian.md`). Earlier entries are not rewritten; each correction below applies from today. Full workings: `system/cycles/001/librarian.md`. The curated canon of the collector's first-hand facts and rulings is now `system/CANON.md`. - **Correction to D116 (the portal ruling).** D116 says "the 'glitch in the portals' puzzle (November 2021) was made by the collector and promoted by Pindar, so it is not a Pindar precedent". Pindar's posts show two different things were merged. (1) The collector's puzzle is "A New Portal" (Pindar, 25 October 2021): "created by bitGANg member Grammar Error", prize "Ghostly spookyGAN (0o0)" (#193), solved by Mazzone11 with an acrostic and a Fibonacci letter shift. This matches the collector's solution sheet and his words "I made the portal glitch puzzle and Pindar kindly promoted it" (7 August 2026); D26 recorded it correctly. (2) "The Final Clue?" (Pindar, 13 November 2021) is a different event: "noohp discovered a glitch in the portals... One frame of each testnet bitGAN reveals mysterious figures, with tentacles", on "Green Shimmer testGAN (3)", "Yellow Shimmer testGAN (9)" and "Violet Shimmer testGAN (27)"; TALK2AM called the numbers "cubes" (they are powers of three) and noticed "81 is next in the sequence and that there were 81 testnet bitGANs in the drops". The post credits nobody but Pindar's testnet pieces with the frames, and names Grammar Error only as one of two people who "got better looks at the glitches". So the November glitch is a Pindar event found by a viewer, and IS-119, SGN-018 and every component that cites D116's ruling should read it that way. Two limits: the carriers were testnet bitGANs, not pieces of the 512; and it is not Pindar's only pre-clue glitch found by watching, because his post "Hidden Types" (31 August 2021) credits thatbloom with the "fibonacci-glitch" seen in the panGAN and TJC345 with the same glitch on grid positions 1, 2, 3, 5, 8, 13, 21. The collector's own reading (26 August 2026): "The testnet precedent I think just references podGANs, which have tentacles." Whether he is Grammar Error, and whether he had any hand in the November frames, is a question for him. - **`pairs.json` count.** Its `_meta` says "count": 11 and "Exactly 11 pairs; #175 sits in two". The file holds fifteen pairs (P01 to P15, 28 distinct pieces). Eleven was true when written (P10 and P11 were added on 6 September 2026); P12 and P13 were added on 20 September (D87), P14 and P15 on 24 September (D105). The collector withdrew P13 (#256 with #479) on 21 September (D91), so fourteen pairs stand. His calls in the file: yes on P01, P02, P03, P04, P07, P14, P15; unsure on P05, P06, P08 to P12; no on P13. #479 also sits in two pairs (P02, P13), not only #175. The file itself is not edited here. - **The Solstice word run skips four words, at four places.** The brief is right. The Solstice drop (#209 to #300, 92 pieces) holds words 129 to 224 of the English list except four: 'awkward' (135, held by #141), 'bean' (157, #21), 'blind' (191, #131) and 'blue' (195, #189). The skips fall after #214, after #235, at #269 and after #271. The card `system/cards/invC_run_break_events.md` counts three skip points because the skip of 'blind' at #269 sits beside the #259/#268 swap and was folded into it. Under its own "one event per cause" rule the forced runs give seven events, not six (#21, #57, four skips, one swap). - **The SpookyGAN prize was #193.** Pindar's post names "Ghostly spookyGAN (0o0)", which is #193 (Type Ghost, a spooky-drop piece in a ghost costume). The collector's words "one of the Ghost costume GANs" (FB, 4 October 2026) describe the same piece: D106 records #193, #196 and #201 as the ghost-costume spookies. So the brief (Part 6, item 4) is right in substance but should name #193 and the date 25 October 2021, so that nobody reads it as a Ghost-drop piece. D118's "awaiting the collector's confirmation" can be closed: his words and the post agree. - **The roboGAN clue was a picture and a data file, not a text.** Already corrected in D121 and FINAL_REPORT; the brief (Part 6, item 2) still says "clue text". Pindar's solution post (28 September 2021) lists "roboGANproperties.csv" and says the twelve words "were the only BIP39 words contained in the clue. (In red below)", then ordered by mint order. - **"Their essence is just a distraction" is a Solstice clue, not an Act 1 clue.** It is the third of the four Solstice clues (journal 014 and the post "Pattern Search": "1: The seed phrase and its order is hidden in the 92 bitGANs of the Solstice Drop ... 3: Their essence is just a distraction, they hide their true BIP39 nature elsewhere."). The clue scorecard (BRIEF Part 2 and PRINCIPLES.md, item 7) lists it among the Act 1 clues. It should be scored as a past-puzzle precedent under item 10, not as a clue to Act 1. The collector said the same on 24 August 2026: "He did not definitively say that the essences were a decoy. He said taht about other puzzles but not this one." - **"All twelve solved roboGAN pieces sit behind the binary" (12 of 12): true, but weak.** In the collector's notes of 26 August 2026 all twelve (#133, #138, #143, #148, #153, #158, #163, #168, #173, #178, #183, #188) are marked as sitting behind the binary ("sits behind binary" and similar wordings). His robo audit of 24 September 2026 says it again for nine of them; for #168, #173 and #183 it describes only the overlay, and on #188 it adds "Gan sits behind overlay (as usual)". Across both files 19 of the 64 robots are marked behind the binary, out of about 24 robots whose notes mention a binary overlay. So sitting behind the binary is the usual look of overlay robots, and it marks the cryptoGAN species at least as much as the solved answer. - **The clue's own seed-list words on pieces** (the lead's script, rechecked today on the sheet and the chain, which agree): 'seed' is the essence of #357, 'random' of #468, 'word' of #183, and no piece holds 'phrase' (nor 'glitch', 'glitches', 'eleven' or 'end'). #183 is one of the twelve pieces of Pindar's solved roboGAN answer. 'final' is #512. - **Test 3 as recorded.** D97 records the collector's third test as "the word 'random' or a truly random checksum-compliant word". The 4 October brief and PRINCIPLES.md state it as "the twelfth is not derived", which drops the first option. His verbatim 21 September review is not in the record. - **D110's "Colour: warm first half, cool second half"** was overruled by the collector's eye on 4 October 2026 (D119): "The review did not have warm/cool halves." - **CANON_FULL.md holds thirteen statements from another project.** `system/build_canon.py` takes every typed message from every session, so statements dated 25 May to 15 July 2026 (sessions 1600968f, d21f56e3, e6494672, about story generation) appear at the end. Nothing before 7 August 2026 concerns bitGANs. The builder should start at 7 August 2026. - Next D-number: D123. ## D123 — The solving system, built 5 October 2026: its ledger, parts, pathways and roles; the collector's rulings and eye answers of the day; and cycle 001's results (1,371,635 phrases, no match) (2026-10-05) Written by the Recorder of cycle 001. Earlier entries are not rewritten. Sources: `system/DESIGN.md`, `ARCHITECTURE.md`, `PATHWAYS.md`, `CANON.md`, `cycles/README.md`, `LEDGER_REPORT.md`, `PARTS_REPORT.md`, `ENUMERATION_REPORT.md`, `cycles/001/*.md`, and the finished job files in `system/done/`. Every count below was read from those files or recomputed by script on 5 October 2026. No odds appear anywhere in this entry, by the collector's ruling of the same day (below). - **Why it was built.** The collector asked for it in these words: "design the system to help me solve this. Not just run a few ideas with a grading tournament ... review/vet past ideas, extend them to exhaustion, learn new strategies, fold them back into the mix, and keep going until we solve this." When the design was ready he said "Build it", and added: "I don't want it to ONLY focus on my large list/cut down strategy. That should be one of multiple pathways. part of the job is to think up viable pathways." The design (`system/DESIGN.md`) names six lessons from the two swarms: the record is bigger than any agent's memory; ideas were tested a few phrases at a time with loose book-keeping (144,000 damaged entries in the old tested list, a coverage bug that hid 4,213 phrases, "never tried" ideas that had run in September); one model graded its own kind of work and the gate kept tightening; testing is cheap (about 90 eleven-word phrases a second) and thinking is the scarce thing; the wallet gives no partial credit, so the system can only learn from Pindar's habits, from what the collector knows and sees, and from what has been covered; and the blind spot is how the pieces get picked. Every hypothesis is now a recipe of named parts: a pool (pieces picked by one trait or mechanism, any size), a reducer (a position rule that cuts the pool), a reader (how a piece gives its word), an optional transform, an order (mint first) and a shape (eleven; ten plus 'final'; a heading plus eleven). Every phrase carries a label naming its recipe, so a hit would name all of its parts. - **The ledger** (`system/ledger.db`, `LEDGER_REPORT.md`). A database of every ordered eleven-word phrase ever tested against the prize wallet, counted only where the run finished and saw its planted test phrase. It holds 2,772,764 ordered phrases from 365 valid jobs (366 with the repair below), 831,438 distinct word sets behind them, and 296,947 legacy word sets whose order nobody can recover (5,876 of them have no valid ordered run at all and are treated as untested). Three quarters of the tested work comes from 68 August job folders in the work folder (2,082,097 phrases), not from `sweeps/` or the swarms; a record built from the swarms alone would have missed it. 24 runs were left out as invalid, unfinished or unreadable (for example `q` and `wq`, 400,000 phrases run in August by a runner with no planted test; `q0827g`, 232,000 phrases with no plant seen). The 4 October swarm's 31,178 phrases were 31,156 distinct, and 121 had already run before it. **The repair of j6:** the August job `j6` had sent at least 36 of its 181 chunks through a second runner that had no planted test, and the chunk files do not say which runner did which chunk. All 140,163 of its distinct phrases were run again on 5 October on the standard path with a plant (`system/done/900_repair_j6.json`: complete, plant seen, no match). A bulk find-and-replace on 28 August had also stamped 181 chunk files in 12 jobs with one false time; the rebuild now dates those jobs from their plant files. The ledger answers three questions before any phrase is built: has this exact order run; have these words run in any order; which jobs ran it. - **The parts library** (`system/parts/`, `PARTS_REPORT.md`, `ENUMERATION_REPORT.md`). On 5 October the registry's check printed: 2,337 pools, 112 reducers, 273 readers, 38 transforms, 30 orders, 6 shapes, 2 layouts, 4 keys, 1 witness and 8 moves; 134 part tests pass (one further test, of an enumerator batch, fails, and it is not a part test). Each part carries a plain description, a reason, the source of the reason and an evidence level: `pindar` (his own words or a solved puzzle of his), `collector` (his recorded eye or statement), `fact` (a scripted fact), `community` (Pindar's 9 January 2022 list) or `analysis` (an agent's inference, the weakest). An independent review found the first labels too generous (too many parts marked as resting on Pindar or the collector); 86 pools and most reducers were relabelled. Among readers, only five rest on Pindar (the essence; the printed counter as a word number from 1; his Solstice date reading in two forms; the SLIP-39 hop), and no solved Pindar puzzle ever flipped, half-turned, shifted or reversed a word: the solved twists were a list hop, a round offset and a reading by the drop's theme. Mint order is the only order with Pindar evidence. Parts read frozen copies of the old matrix registry, the Phase 1 modules and the site files (`system/vendor/`, each with its checksum), never the live swarm folder. The enumerator ran the collector's Generate, Reduce, Order strategy across the whole library: 1,231,072 pairings of pool and reducer setting (111 reducers in 527 settings), giving 5,742 different elevens, 6,806 tens (read as ten plus 'final') and 2,349 twelves with one member singled out by a stated rule. Its first batch (661 phrases in three rings) ran in cycle 001. - **The 41 pathways** (`system/PATHWAYS.md`, one card each in `system/cards/`). A pathway is a whole way of attacking the puzzle: what counts as a glitch, how pieces get picked, how words come out. No pathway is favoured by default. PW01 to PW12 were seeded from what the project had already done (one trait exactly eleven; generate, reduce, order; five drops solve the sixth; every drop holds a portion; Ghost first; keys and locks; the Review as a map; one step away; Pindar's own list; a glitch is a place; a pointer outside the trait table; the collector's eye as the selector). Five Inventors, on two models, added nineteen (PW13 to PW31: invC on Fable, from 'seed' to 'random', run breaks, a passphrase, a longer phrase, the collection as the seed itself; invD on Opus, portal frames, eyes in the background, speed by SubType, what the Review changed, behind the binary; invB on Opus, the sixteen-piece block code, each kind of glitch locates one piece, the Ghost grid without its 'ex' cells, find the other planted columns, the planted blocks as labels; invA on Fable, the number spiral, there and back, the hidden types as one hex number, each bitGAN is one bit) and 50 strategy cards in all. **The collector added five, in his words** (PW32 to PW36): "chain/arweave csv differences"; "visual vs. trait discrepancies"; "glitches as describing moves (swap, reverse, delete, etc)"; "description differences leading to a different arweave sheet (in Ghost Drop I believe)"; and "we also need to cover 'themes' of the drop (colors, fibonacci, palindromes, etc). These are different than the mathematics of the sets, though there may be overlap." On PW34 he added: "for Glitches as moves, consider that this may apply to a much larger set found through a different mechanism, or that it may apply contingently based on the drops in which they occur." Inventor invE (Fable) then added five under PW36 (PW37 to PW41: ghosts in the machine read as a rule; the Review answers in opposites; every drop is partly a review; the six themes there and back; the story's chapter titles). An Excavator (Opus) wrote cards for PW32 to PW35 and found, among other things, that `primary_facts.py`'s chain-wrong list wrongly holds #309 (the sheet is the odd side there) and that the one description leading to a different Arweave sheet belongs to Solstice #272, not a Ghost piece. - **The roles and the model rule** (`system/roles/`, `cycles/README.md`). Librarian (rebuilds the record and the canon); Inventors (new part types and strategies, judged on being new and buildable, not on plausibility; several at once from different starting points and on different models); Excavators (turn past ideas into parts, checking what exactly was tested and whether it was built right); Builders (code and test parts; a part enters only when its test passes and a second model has reviewed it); Assemblers (compose whole hypotheses from parts); the Grader (stands in for the collector, checks that an idea is real, reasoned and correctly built, reruns the lookups behind its premises, sets priority; his principles are weights, not pass marks); the Auditor (every third cycle, audits the grader, the directions, the coverage numbers and the book-keeping, and backtests new part types against Pindar's solved puzzles); the Reframer (the step between cycles, below); the Recorder (this entry); and the Liaison (the one voice to the collector: a short digest and a live-GIF review page for the few questions only his eye can answer). The model rule, his words: "rotate between Fable and Opus in reviewing the work. I don't always want the same model grading the work of itself on other tasks." So reviewers alternate, and a model never grades its own work. In cycle 001 the Assemblers composed on both models and the Grader was Fable 5.1 (it rebuilt every Fable-composed phrase with its own code rather than sampling); the Reframer was Opus 5.5. - **The research rule and the record index** (`kb/RESEARCH_RULE.md`, `kb/lookup.py`). His words: "when they propose something, they should research all aspects of that proposal from teh full data set before they full form their theory." Every proposal, report and review now carries a "Premises checked" section: each premise, the lookup run, what the record says including what cuts against it, and a verdict. The index holds 20,112 passages (the collector's own words 1,312; primary facts 4,162; decision entries 149; Pindar's texts 72; agents' analysis 14,417), ranked by authority in that order. `system/CANON_FULL.md` holds his 363 typed statements from every session (the Librarian found thirteen from another project, dated before 7 August 2026, which are ignored), and `system/CANON.md` is the curated canon every agent reads first: his first-hand facts, what his eye has settled piece by piece, his three tests and principles, his rulings on how to work, his own ideas, and what he says he does not know. - **His rulings of 5 October 2026** (verbatim, from `CANON.md` and `ARCHITECTURE.md`). - No odds: "I would avoid calculating odds. Who knows why Pindar chooses what he does or how likely the moves are. You might get a lot of false negatives there (false positives are fine because at least we check)." So no idea is dropped, held or ranked down because a pattern is at chance or was found after looking; the chance desk is retired as a gate and as a weight, and odds are not reported unless he asks. - Depth by default: "I'm finding that you are throwing WAY too much info back at me ... I'm worried that I'm missing things -- particularly opportunities to ask you to delve deeper on promising but incomplete leads. And that we are still looking at so few phrases per idea. I feel like we need to run more permutations based on library elements." So every selection an idea produces is crossed with every reader that gives a word for all its members, every transform with a stated reason, and every distinct order, skipping what the ledger holds; permutation is over library parts only, never substitution of arbitrary words or pieces. - The Reframer step: "We also need to have a step between rounds that consider ways in which a theory that did not work may have been overly constrainted/narrow, may be overly broad, or may be just a piece of the puzzle used in a different way. We cannot fully discard things when we may just not yet know how to use them properly." The step is `system/roles/reframer.md`; for every missed hypothesis it asks whether it was too narrow, too broad, or a piece used the wrong way, and writes each reframing as an exact recipe with its support and the nearest thing already tested. It ran for cycle 001 (`cycles/001/reframe_reframe.md`, Opus 5.5): 17 records and twelve ranked recipes for cycle 002. - Visual reviews: "Anything visual you want me to review must be presented as an artifact with live GIFs." Also, on the Ghost grid: "I don't want to review all the GANs in the layout. I want YOU and your swarm to search for patterns along the rows." - Pathways, brute force and the gate: "I don't want it to ONLY focus on my large list/cut down strategy. That should be one of multiple pathways." "brute force should be used with care otherwise it's just an endless unguided use of resources." "I agree we only keep the standard wallet path." "Runs can go for hours/days with the ability to pause". On the audit of the swarm's grader: "Yeah, those bars were too strict for the grades." On results: "If they start coming back with nothing new consistently, I want you to evaluate and come up with a completely new strategy." On drops: "I don't think that needs to be an equal number of GANs from each drop. But I do think the puzzle probably needs to include at least the Ghost Drop and Act 1 Review. If a rule works for more drops = more credibility in the theory, not a hard bar that has to be met." And on the blocks: "I've very bullish about these ex/so blocks and finding 11's inside them or across drops. We need to exhaustively examine that." - **The corrections in D122** (the Librarian's, same day) stand and are applied in the system: the November 2021 "glitch in the portals" is a Pindar event found by a viewer (noohp), on testnet pieces, and is not the collector's puzzle; the collector's own puzzle is "A New Portal" (25 October 2021), prize #193; `pairs.json` holds fifteen pairs, fourteen standing; the Solstice word run skips four words at four places; the roboGAN clue was a picture and a data file; "their essence is just a distraction" is a Solstice clue, not an Act 1 clue; "behind the binary" is the usual look of overlay robots; the clue's own seed words sit on #357 'seed', #468 'random', #183 'word' and #512 'final'; and test 3 as recorded allows the word 'random' as the twelfth. - **His eye answers of 5 October 2026** (saved on the live-GIF review page; verbatim in `CANON.md`). On #430, asked which label it should carry: "It's like a glowing/ultra hybrid. Maybe that is intentional as the pieces move through the colorwheel in this drop." On #339 with #340 and #429 with #430: "Yes, both pairs look swapped." On #145 and #150: "Yes. Each looks like it carries the other's label." On the Ghost drop's first column (the twelve pieces at the left edge of the grid laid twelve wide): "No discernable visual pattern here". These close the three questions D121 left open. The jobs that assumed the two Ghost swaps (five_solve_sixth_r5_1; recover_a_r6_1's Ultra and Vivid by eye; recover_b_r6_3) ran the same day and did not open the wallet. The lead's note: there are four Shy pieces, #338, #340, #355 and #357; his earlier review said "Shy: all mostly yellow except for the spaceGAN which is more green", and #340 is that spaceGAN. - **What ran on 5 October 2026, before the cycle** (the closure engine, standard path only, each phrase with all 128 valid twelfth words, plant seen, no match): audit_p1, 1,530 phrases, the seven held files the outside audit (`GRADER_AUDIT.md`, Fable) said should have run; recovery_r6, 51 phrases in nineteen jobs, the recovery round under the reset gate (graded by Fable 5.1); borrowed_r6, 30 phrases in four jobs, Pindar's 9 January 2022 list of community theories carried to Act 1 (graded by Opus 5.5); labelwalk_r6, 12 phrases, Pindar's theory 11 read as he wrote it (every Type and SubType label that is a seed word, walked down a drop with repeats dropped, which gives exactly eleven in the Ghost drop and exactly eleven in the Review from #445, and fewer in every earlier drop). 1,623 phrases in all; the ledger later found four of them had run in August on 33 paths. - **Cycle 001's results** (`system/done/*.json`; every file says complete, plant seen, no match; standard path; each phrase with all 128 valid twelfth words). One card per family went to the site. - The one-phrase test run ahead of grading: c001_asm_words_1, the 'ex' and 'so' block pieces inside the window from #357 'seed' to #468 'random' (seven 'ex' and four 'so' pieces), by essence in mint order. 1 phrase. - The graded ideas: 35 assembled jobs (asm_words 15 jobs and 519 phrases; asm_themes 6 and 68; asm_revisit 9 and 49; asm_grids_eye 5 and 151) and the enumerator's first batch in three rings (138, 83 and 440). 1,448 phrases. The Grader held nothing on odds and sent three things back for an exact change (one job whose eleventh member had no source in his notes; one phrase with two counting bases in one rule; two tasks with an unstated wrap). - The ex/so study (Opus 5.5): eleven ideas on the two planted blocks, 107 selections, each with the full depth pass. 447,197 phrases. **The verified "third family" fact:** turned over on the list (word n becomes word 2,049 minus n), the 'ex' range 627 'exact' to 649 'extra' lands on words 1400 'puzzle' to 1422 'random', and the column's head 'exact' becomes 'random', the clue's own last word; nine pieces in three drops hold a word there: #153 puzzle, #317 pyramid, #130 quantum, #367 question, #454 quick, #170 quiz, #438 race, #134 rain and #468 random (robo 4, Ghost 3, Review 2). Five hold the turned-over form of a word the column holds and four of an 'ex' word no piece holds; the 'so' block reaches the same window by 256 back along the list. Checked by script on 5 October. Its words did not open the wallet in the orders tried; the set stays on file as a cross-drop pool, and whether these pieces look like a set to his eye is a question for a live-GIF page. The study's scan for other planted prefix blocks found none of the same kind: outside Pindar's known index-column runs, only the 'ex' column, the 'so' lines, and runs of three ('co' at #372, #380, #389; 'co' at #319, #344, #368; 'tr' at #455, #465, #474). - The layout atlas and the collector's four row directions (Fable 5.1): ten ideas, 3,580 sets. 761,366 phrases. **The atlas's plain results:** on the Ghost grid laid twelve wide, no trait, glitch or type has one member in every row; only the twelve columns and the 'ex' words do; no field names an odd one out in every row; 'extra' read as "ignore row twelve", the slip of the last four 'ex' words read as a shift (of the cell, the piece number and the word), and every layout where the line does not break were all built and run; removing any one of #386 to #397 straightens the 'ex' column and the collection names four candidates for the 'extra' piece (#389 'correct', the word at its own number; #390 'century', the one piece with two chain records; #393 'twice'; #397 'bullet', the cell the slip passes over), while inserting a blank never straightens it; across every drop at every width, the scan re-found only the 'ex' column, Pindar's solved roboGAN answer (every fifth robot) and the six 'so' words in Review column 7 with one slip. Not built, for the lead: blocks and checkerboards outside the Ghost and Review layouts (about 12,000 sets) and the full pass on counting-rule lines. - The depth pass (the lead's script over the 232 selections the Grader verified): three files finished, 60,000 phrases; the rest were in the queue when this was written. - The exhaustive ex/so subsets pass the collector asked for (every eleven of the 24 block pieces by essence in mint order, 2,496,144 in 25 files): file 1 finished, 100,000 phrases; files 2 to 25 were in the queue. - In all on 5 October: 1,371,635 new ordered phrases (1,623 closure plus 1,370,012 through the system's queue), plus the 140,163 j6 phrases run again. The planted test phrase was seen in every job. No match. Under P18 each miss means that exact combination did not open the wallet; every part stays available. - **Site records with this entry:** 10 card entries appended to `site_sync_pending.json` (74 to 84): audit_p1, recovery_r6, borrowed_r6, labelwalk_r6, the one-phrase seed-to-random job, and one card each for the graded ideas, the ex/so study, the atlas, the depth pass so far and the exhaustive pass so far; for the one approved sync command, card numbers assigned when it runs. 24 per-piece records appended to `theories/anomaly_ledger.json` with source "system_2026-10-05" (4,339 to 4,363): the nine third-family pieces; the four candidates for the 'extra' piece in the 'ex' column; his eye answers of 5 October on #430, #339, #340, #429, #430, #145 and #150 as his own observations; and the four Shy pieces. Tags: #457 added to the "Colours cycle faster than normal" tag in `tag_schema.py` and `bitganspuzzlebook/scripts/build_tags.py` with a dated note (his "Yes 457 speeds up" of 4 October; the tag had been built from his notes before that); two tags added, "The 'ex' column turned over: the nine pieces from 'puzzle' to 'random'" (data, checked by script) and "Label swapped with its partner's, by the collector's eye (5 Oct 2026)" (#145, #150, #339, #340, #429, #430). `build_tags.py` was read first (local-only: it writes `db/tags.sql` and one JSON file, no network or deploy step) and run: 533 tags (531 before), 9,621 assignments (9,599 before; the 22 new ones are the sixteen new members plus #457's six theme roll-ups). Nothing was synced or deployed. Details: `system/cycles/001/site_record.md`. - Next D-number: D124. ## D124 — Cycle 002 of the solving system, so far: the Builder's thirteen parts, the Reframer's twelve recipes (63,248 phrases, no match), the six leads the collector approved and tier C (3,025,796 phrases queued), and a correction to `drop_rules.json` (2026-10-06) Written by the Recorder of cycle 002 (part 1). Earlier entries are not rewritten. Sources: `system/cycles/002/build_build_002.md`, `assembly_asm_reframe.md`, `grading_reframe.md`, `assembly_asm_leads.md`, `grading_leads.md` (with its tier C section), the `hyp_asm_reframe_*.md` and `hyp_asm_leads_*.md` files, the finished job files in `system/done/`, the lines in `system/queue.log`, the files in `system/queue/`, and the collector's words in `system/CANON.md` and `theories/swarm_2026-10-04/COLLECTOR_FEEDBACK_2026-10-04.md`. Every count below was read from those files or counted by script on 6 October 2026. No odds appear anywhere in this entry (the collector's ruling of 5 October: "I would avoid calculating odds"). Every miss below means only this: that exact combination did not open the wallet; its parts stay available. - **The collector's words that drove the cycle** (all 5 October 2026, verbatim). On the blocks: "I've very bullish about these ex/so blocks and finding 11's inside them or across drops. We need to exhaustively examine that." On the Ghost grid: "I want YOU and your swarm to search for patterns along the rows. Maybe there is a certain type in each line, or a certain glitch in each line. Extra may actually be 'extra' - as in ignore that line. The shift of the last four over one space may indicate you need to do a shift. Perhaps there is a visual way to lay them out where the line between them doesn't break, etc." On depth: "we are still looking at so few phrases per idea. I feel like we need to run more permutations based on library elements." On missed theories, which is why the Reframer step exists: "We cannot fully discard things when we may just not yet know how to use them properly." And at night, in reply to the six-lead pick-list and the request to update the site: "all approved. Update the site and keep working overnight". His earlier strategy, which lead 6 applies, is in his words too: "much larger lists than 11 but then applying a second filter like Fibonacci placements". - **The Builder's thirteen parts** (`build_build_002.md`; Opus 5.5, tag `build_002`, code in `parts/cycle002.py`). The Reframer of cycle 001 asked for thirteen parts and the lead asked for three layouts. Built: two held-partner readers (an 'ex' word read through the word 1,024 places on, or through its turned-over word, where a piece holds it); the keys `ex_places`, `so_places` and `eye_robot_counters` with a reducer that keeps the pieces at a key's places; a reader `column_ex_word`; the transform `slip_last` and the order `rotate_last` (his "shift of the last four"); the order and layout with #390 at its first chain record (slot 519); a reducer that cuts a twelve into its eight and its four; the swap pools (the firm ten: #7/#8, #145/#150, #259/#268 and his two Ghost pairs; and all twelve with #487/#494) and a transform that undoes the swaps; a shape that reads the first word twice (the repeated 1 of the Fibonacci list) and the 112-piece 'seed'-to-'random' pools; an order `mint_from`; #430 "as Ultra, as Glowing, or both" built as three pools and an order, because a label changes no word; an order `turn_at`; and the pool `std.flip_holders_9` (the nine pieces from cycle 001's third-family fact). Two requests were partly built or not built, with reasons: the flipped robo answer as places (ten of its twelve numbers fall past the end of the joint release, and nobody has stated a wrap rule), and a stand-in reader for #473 'FALSE' (not needed: 'false' is word 659 of the list, and the essence reader already gives it). The three layouts are the joint release #301 to #512 twelve wide, the Ghost drop in mint order with #390 at slot 519, and the Ghost rows each starting at its 'ex' piece. The Builder also repaired the one failing registry test (the test, not the batch, was wrong: the batch had since run). The registry check then passed 141 tests, and the library held 2,425 pools, 115 reducers, 276 readers, 40 transforms, 35 orders, 7 shapes, 5 layouts and 8 keys. - **The Reframer's twelve recipes and their results** (`assembly_asm_reframe.md`, `grading_reframe.md`). The Reframer (Opus 5.5) looked again at every idea that missed in cycle 001. An Assembler on the other model (Fable 5.1) built all twelve recipes, each with the full depth pass, and a Grader on the first model (Opus 5.5) rebuilt 10,670 phrases with its own code (no mismatch), repaired labels on three jobs (words untouched, hash-checked) and approved all twelve. All twelve ran on 5 October (`system/done/140` to `180`), each complete, plant seen, standard path, each phrase with all 128 valid twelfth words, no match. By pathway: PW05, 41,425 phrases (the 'ex' column through its held partners, 608; the joint release twelve wide with the slipped line followed into the Review, 3,115; the ex/so pass continued in the ten-plus-'final' shape, 20,000; the Ghost grid with #390 at slot 519, 17,702); PW02, 7,185 (the 112-piece stretch from #357 'seed' to #468 'random' at Fibonacci places); PW06, 5,699 (block places as keys across the two drops, 1,580; the swap twelve in five drops with 'sort' as the instruction, 3,248; the robot counters of his two eye groups as Ghost places, 871); PW34, 6,906 (the planted-error program, 2,007; the program from his on-purpose list D90, reverse then shift, 1,959; the slip as a move on the last four words, 2,940); PW36, 2,033 (#430 as a Glowing and Ultra hybrid, with the Ghost swaps undone). Total 63,248 phrases. The Assembler's three corrections to the Reframer's record were confirmed by the Grader's script: #473's 'FALSE' is a seed word (word 659); the straight 'ex' pieces have six 'so' partners, not four; and `token_index.json` stores slot 583 for #390, not 519 (a dated correction record is now on #390 in the anomaly ledger; the earlier record stays). The Grader's notes for later: the real slip moves pieces, not words (built in lead 2 below); three readers added after the build deserve a short top-up pass; and the record says Pindar's "112 ghostGANS" is the Ghost drop less its 32 testGANs, not this stretch, with the post dated after the prize went in. - **The six approved leads, built and queued** (`assembly_asm_leads.md`, `grading_leads.md`). After the collector's "all approved", an Assembler (Opus 5.5, `asm_leads`) built every reasoned form of each lead, each with a full depth pass; a Grader on the other model (Fable 5.1) rebuilt 5,746 phrases (no mismatch), found none of 3,099 sampled phrases in the ledger and none of the 1,875,446 in the queue, the finished jobs or the reframe files, repaired two label defects (words untouched), and queued all 99 files: tier A (the forms the leads name) at 400 to 489 and tier B (cuts of pools the Assembler put together, with reducers whose reason is Pindar's or the collector's) at 700 to 734. The six leads, with tier A and tier B counts: (1) the third family, the 'ex' words turned over, read with the pieces that hold them, and the same for the 'so' block (PW08, PW06; 25,428 and 245,814); (2) eight plus four, both blocks split as the 'ex' column's slip splits it, including the slip read as a shift of pieces (PW05, PW06, PW25; 24,570 and 209,957); (3) the 'extra' piece, each of the four candidates #389, #390, #393 and #397 set aside, and the straightened grid read along every line (PW25, PW26; 149,700, Ghost only); (4) the swaps, his three by eye with Pindar's two and the Review pair, as pieces, as moves and as keys (PW33, PW34, PW06; 62,131, five drops); (5) matching rows, the six 1,024 pairs used to line up the Ghost and Review grids row by row (PW06; 61,355 and 98,896); (6) the 112-piece stretch under every filter but Fibonacci (PW13, PW02, PW06; 997,595). Totals: tier A 1,320,779, tier B 554,667, together 1,875,446 phrases in 99 files. How his words map onto them: "very bullish" on the blocks (leads 1, 2 and 5); the search along the rows (leads 3 and 5); 'extra' as "ignore that line" (lead 3); the slip as a shift (lead 2); his "much larger lists than 11 but then applying a second filter" (lead 6). None of these files had run when this entry was written. - **Tier C** (end of `grading_leads.md`). The Grader asked for the reducers left out of tier B (those whose reason is a scripted fact or the library's own analysis: palindrome and class places, every k-th from 2 to 11, the first ten to twelve, the two ends, the mirror runs) on the same assembled pools, citing his "false positives are fine because at least we check" and "we need to run more permutations based on library elements". The lead (Fable 5.1) built them on 6 October with `cycles/002/build_asm_leads_tierC.py`, skipping every phrase already queued, finished or in the ledger: lead 1, 721,661 phrases in 37 files; lead 2, 47,093 in 3 files; lead 5, 381,596 in 20 files; 1,150,350 phrases in 60 files at 800 to 864. Leads 3, 4 and 6 had no deferred pools. None had run when this entry was written. - **The correction of 6 October 2026 to `drop_rules.json`.** The Solstice rule "The essences run the wordlist in order and the mint order equals the piece order" named its break as "#259 'hungry'". The chain essence and the sheet essence of #259 are both 'bless' (checked by script from `primary_facts.py`), and no piece in the collection has the essence 'hungry'. The line now reads "#259 'bless' jumps ahead of the run and #268 'bird' falls behind it", with a dated note of the correction kept in the file. - **What else finished since D123** (the cycle 001 work still in the queue then). The last two depth files, `c001_depth_04` (20,000) and `c001_depth_05` (6,952), finished the cycle 001 depth pass: five files, 86,952 phrases. Files 02 to 07 of the exhaustive ex/so pass (every eleven of the 24 block pieces by essence in mint order) ran, 100,000 phrases each. The pass holds 2,496,144 sets; the 25 files hold 2,496,119 phrases, because the build skipped sets the ledger already held (file 25 holds 96,119). Every job complete, plant seen, standard path, no match. - **New per-piece facts** (anomaly ledger, recomputed by script from `primary_facts.py` and the word list): the three pieces that hold a turned-over 'so' word (#342 crack from 'soda', #344 couple from 'solution', #389 correct from 'sorry'), with the 'so' eleven that follows (the ten unpaired 'so' pieces plus #342) and the fact that the twelve 'so' words turned over give list numbers 383 to 394, a run of twelve Ghost pieces around #389; the five 'ex' pieces whose turned-over word a piece holds (#301, #385, #398, #422, #434) and the two 'so' pieces that have one (#450, #463); the six 1,024 pairs (#325/#450, #337/#479, #349/#487, #361/#499, #373/#458, #385/#502), two of them in matching rows with both 'so' pieces in Review column 7; and the correction on #390. 23 records on 21 pieces. No tag was added. - **Totals so far, counted by script on 6 October 2026.** Run since D123: 690,200 phrases (cycle 002's twelve reframe jobs, 63,248; the last two cycle 001 depth files, 26,952; exhaustive ex/so files 02 to 07, 600,000), every one with its plant seen and no match. Waiting in `system/queue/`: 4,821,915 phrases in 177 files (the ex/so pass files 08 to 25, 1,796,119, with file 08 running; the leads' tier A, 1,320,779 in 70 files; tier B, 554,667 in 29 files; tier C, 1,150,350 in 60 files). The ledger's phrase table, read only, held 4,832,962 ordered phrases from 506 valid jobs (2,772,764 when the system was built on 5 October). Nothing has opened the wallet. Site cards for the finished jobs (twelve cards: five for the reframe pathways, one for the depth files, six for the ex/so files) are in `site_sync_pending.json` for the lead to sync. Details: `system/cycles/002/site_record_part1.md`. - Next D-number: D125. ## D125 — Cycle 003 of the solving system, so far: the audit and its sixteen fixes, the collector's rulings of 8 and 9 October, eight new pathways (PW42 to PW49), the Builders' new parts, the Excavator's findings, his eye answers of 9 October, and 5,255,990 more phrases run (no match) (2026-10-09) Written by the Recorder of cycle 003. Earlier entries are not rewritten. Sources: `system/cycles/003/audit.md`, `build_build_003a.md` to `build_build_003d.md` with `review_build_003a.md` to `review_build_003c.md`, `excavation_exc_lenses.md`, `assembly_asm_drops.md`, `assembly_asm_empty.md`, `assembly_asm_mirror.md`, `assembly_asm_eye.md` and their four gradings (`grading_drops.md`, `grading_empty.md`, `grading_mirror.md`, `grading_eye.md`), `system/CANON.md` (sections "8 October 2026" and "9 October 2026"), `system/PATHWAYS.md`, `lore_catalog_runs/bitgang1001/frame_match_2026-10-09.md`, `system/cards/invF_channel_map.md`, `system/queue.log` and the finished jobs in `system/done/`. - **The audit (Auditor Opus 5.5, 6 October 2026; `audit.md`).** The books are clean: every run so far (141 jobs, 2,200,374 phrases at the time) finished with its test phrase seen and no match, and the ledger, the run log and the finished files agree to the phrase. The graders had not brought the old bars back, with one mistake: the cycle 001 grader held the portal-frame job (PW18) because "the #357 description is not in the record", but it is in the record (decision D4b, 27 August 2026), and that job and two small send-backs were never rebuilt. The bars had moved upstream: in cycle 001 the Assemblers left ideas unbuilt because a pattern was "at chance" or a count came out at twelve, the clearest case being PW41, the chapter titles of Pindar's posts, which is the pathway most like the one puzzle of his that the community solved (the roboGAN puzzle). Thirteen of the 29 newer pathways had never had a single phrase run, and the collector's own ideas of 4 October ("the 5 previous drops solve the 6th" and "every drop holds a portion") had one phrase and none. Nearly all of the 4.8 million queued phrases came from one family, the ex/so blocks and the 'seed'-to-'random' stretch. Three system problems: only 29 of 1,854 sampled labels could be rebuilt into a recipe by code; the record search ranked Pindar's own posts as the lowest authority; and five groups of parts claimed stronger evidence than they had. The parts themselves were sound (eleven pools and six reducers rebuilt with the Auditor's own code all matched). - **The sixteen fixes of the audit, and where each stands on 9 October.** (1) Correct the #357 hold and run the held work: the six-phrase job ran as `c001_grids_eye_5_by_D4b`, and the wider portal-frame forms were built and run (PW18). (2) Ledger read paths read only: done (`ledger.connect_ro()`; `ledger.py stats` now runs while the runner is live). (3) "One drop" is judged by the drop whose rule picked the pieces: written into `roles/grader.md`. (4) The grader states three weights, not four (chance was retired as a gate on 5 October): written into `roles/grader.md`. (5) The Assemblers' not-built lists are visible to the grader, and "at chance", "below the bar" and "the count is not eleven" are not reasons: written into both role files; the builds owed (the PW23 data reading, the Polybius reading, PW41) were built. (6) Two-move labels and a pool id for the D90 list: done. (7) Tier C graded by Opus before the runner reaches it: done (`cycles/002/grading_tierC.md`, 8 October; Tier C runs as queued). (8) Make labels rebuildable: the first step only, a census test that reports the share of labels that rebuild; the full registration of every label's parts is still owed. (9) The depth pass: the shape `ten_random` is mapped and a dated note in ARCHITECTURE.md states what the depth pass is. (10) `python3 parts/tests.py` runs: done. (11) Build PW41: built (see the results below). (12) An Assembler for PW03 and PW04: done (`asm_drops`). (13) The queue's make-up in every digest: written into the Recorder and Liaison role file. (14) The record index's authority labels: Pindar's own posts are now filed as his. (15) Five groups of evidence levels lowered, each with a dated note on its card: done. (16) The Solstice shape (a reader that reads one field down a selection in mint order and drops repeats, the Winter Solstice answer): built by Builder build_003b as `read_down_unique`. - **The collector's rulings and answers of 8 and 9 October, in his words (CANON.md).** On 8 October, asked (audit fix 13) whether the ex/so family should keep nearly all the machine time or whether a share should go to the pathways that had no phrases yet, he said: "split the pathways". The lead applied this so that the queue alternates between the ex/so family (the exhaustive files at priority 300 and the leads' Tier A at 400 to 489) and new jobs on the empty pathways, which take priorities 500 to 699, ahead of Tier B (700) and Tier C (800); each grading interleaves the two families block by block. On 9 October, from the live-GIF eye page (https://claude.ai/artifact/WbEHiAESibzPdUfpEexhAy, saved 11:33 UTC), his answers are recorded under the heading below. He also agreed to a download: "You can download WordNet." (he first asked what WordNet was). Princeton WordNet 3.1, the dictionary files, 16 MB, is in `external/wordnet/dict`; the opposites pathway (PW38) has not yet been rebuilt on it. - **His eye answers of 9 October 2026, verbatim.** Behind or in front of the binary: #459 behind, #460 behind, #462 behind, #463 behind, #464 behind; #508 "I cannot tell"; every Ghost piece he was asked about (#303, #309, #327, #341, #363, #382, #391, #405, #444) in front. His note: "In the Cyborgs, teh binary is the eyes/mouth. It's part of the face. In Bot Rainbow, that is one of the abstract/corrupt GANs so there is no consistency." Portal-frame targets for #129, #198, #200, #201, #202, #203: none named. His note: "I can't tell. There are GANs hiddne under the costumes but it's not possible to say what they are. I would guess that the traits/metadata might give some indication but visually it's not possible to see." The 32 faces of Pindar's bitGANg 1001 card: none named. His note: "Your cut of the GIF crosses between faces and isn't clean. But even if it was, I don't think I could do a better job than the machine comparing frame by frame. That's the best way to do this." The spelling is his, kept as written. These became per-piece records in the anomaly ledger and one tag (`behind_binary_by_eye`, the five pieces he marked behind). - **The efficiency change (made on 8 October and signed by the Librarian; recorded here for the next cycle).** `system/START_HERE.md` (written on the evening of 8 October) is now a briefing pack for every role: the puzzle in four lines, the hard rules, the recipe grammar, the collector's rulings in his words, the key verified facts and the tools. The first line of every role file tells the agent to read it first and to open the full ARCHITECTURE, DESIGN, PATHWAYS and CANON files only for the sections the task needs. Its Efficiency rules are: read only the files the task names; look things up with `lookup.py` instead of reading whole reports; summarise with python instead of printing JSON; keep tool calls under 40; and write the output file early and extend it. `precheck.py` is the mechanical half of grading (shape, words on the list, repeats, overlap with the queue, the ledger sample, priority collisions). The record index now ranks Pindar's own posts as his (audit fix 14). - **The new pathways, PW42 to PW49 (`PATHWAYS.md`).** Inventor invF (Opus 5.5, 6 October): PW42, Pindar's round offset on the file numbers (the SuperRare reading, a number less a round ten-thousand is a word number, run on what each file measures: 21 pieces land by file size and 4 by duration, #129, #146, #479 and #480); PW43, pieces that are out of step with their drop in time (nine saved outside their drop's working days, six released from Pindar's wallet at a different time, less the four tributes the clue page explains: eleven across Solstice, Ghost and Review, holding 'random' #468); PW44, the Ghosts in the Machine card, bitGANg 1001 (the hidden social card of the clue post, saved minutes before the post, named for the Ghost drop, with 32 faces in a grid); PW45, the hidden Photoshop fingerprints as seed bits (each GIF's 32-hex fingerprint is 128 bits). Inventor invG (Fable 5.1, 8 October): PW46, the Review played backwards over the Ghost drop (piece n with piece 813 minus n); PW47, a 'so' number picks the cell in each Ghost row; PW48, the move words of the joint release as a program, with a planted twelve as the tape; PW49, the three runs of four in the Review that answer the 'ex' column's three species. The pathways now number 49. Pathways with no phrases yet, by the lead's digest of 8 October: PW15, PW16 and PW17 (they need engine changes), PW20 (needs a speed measure), PW21 (needs the collector's side-by-side ticks) and PW24 (needs a syndrome reducer). - **The Builders' new parts.** Build_003a (Fable 5.1; reviewed by Opus 5.5): no new part, seven system fixes (audit fixes 2, 6, 8 first step, 9, 10, 14 and 15; the reviewer found one serious problem in the depth-pass rerun instruction, which would have logged 20,000 untested phrases as tested, and fixed it). Build_003b (Opus 5.5; reviewed by Fable 5.1, who rebuilt a sample with his own code and found the code correct): 12 orders (by save time, upload time, first transfer out of Pindar's wallet, sheet title A to Z, and printed counter, each forwards and reversed, plus Pindar's 39 save sessions), 2 shapes, 1 reducer, 4 keys, 27 readers, 189 single-piece pools and one pool of pairs, plus the Solstice shape. Build_003c (Fable 5.1, finished by a second Builder; reviewed by Opus 5.5): 72 parts for the mirror and its companions (41 pools, 1 reducer, 18 readers, 1 order, 1 layout, 8 keys, 2 moves). Build_003d (four small items from the review of 003c, checked by the grader of the mirror batch): the continued count in the card's two forms (2 keys and 8 readers), the mint-order mirror (1 layout, 1 reader, 9 pools; in mint order only #301 and #512 keep the partner they have in the 813 mirror), the depth selections registered as pools with their evidence taken from the source, and 54 of the 59 September pools lowered from `pindar` to `fact` (5 keep `pindar`: Pindar's own words or his solved puzzle). The registry, run on 9 October, counts 2,898 pools, 117 reducers, 330 readers, 40 transforms, 48 orders, 10 shapes, 7 layouts, 22 keys, 1 witness and 10 moves; 161 tests pass (CHECK PASS). - **The Excavator's findings and errors (`excavation_exc_lenses.md`; Opus 5.5, 6 October).** The slice was the 27 to 28 September swarm: 165 job files, 164 run, 233,389 phrases. What the swarm tested: almost every job crossed one selection with the same seven word rules and the same seven orders, so a single eleven got a median of 28 phrases and at most about 100, and nothing in that slice had the depth pass of 5 October. Four of the seven standard orders were not in the library, and neither were "every eleven of twelve" and "ten plus final" (the reducer `drop_less`). By estimate about 1,336,766 reader-and-order phrases are left, and about 34.7 million if every transform is crossed in. Selections that changed since September: the colour-flash pool is now fourteen, the swap-pair list is five pairs, and the speed-up pool exists only since 4 October. Seventeen errors were found in the swarm's own records, all recorded and none deleted: #30 and #29 swapped mint slots (lens2_02); two formatting differences listed as contradictions, #309 and #383 (markers_r5_3); a soft class edge for the counting words (channel_r1_3); twists_r1_6 was not comprehensive, as the collector had asked; the first-return frame is 37 on only eight robo files, not on all (visual_r2_2); a wrong negative-frame list for the Solstice files and #479 in four jobs (the true counts are six on each Solstice file and seven on #479); #193 counted as a palindrome against his ruling of 13 September ("spookyGANs don't have palindromes. That was only for Robo Drop.") in four jobs, and one pattern that exists only if the F is an error; #119 read as F where Pindar's January 2022 table gives 'd'; forge partners that Pindar's own "Burned bitGANs" trait contradicts (unlock_r6_2); and the claim of thirteen colour words, where the list holds 28 colour-like words and fifteen pieces hold one as their essence (channel_r1_1). Its five leads were built by the Assembler of the mirror batch (see the results). - **The bitGANg 1001 frame match (9 October; `lore_catalog_runs/bitgang1001/frame_match_2026-10-09.md`).** The collector said the earlier cut crossed between faces, and the check confirmed it: the real tiles are about 97 pixels wide on a 105-pixel pitch, not equal cells of 90. Each of the 32 inner faces was cut between the white gutters and compared frame by frame (12 saved grid frames, 55 frame offsets, nine views including grey, mirrored and hue-turned) with all 512 pieces. Result: no face is an Act 1 piece. No face is an exact copy of a piece (the best scores run from 0.025 to 0.070, average 0.043, where a true copy scores about 0.0005 and an unrelated pair about 0.098); the grid frames are not locked to the piece frame numbers; 31 of the 32 closest pieces are Ghost pieces (the other, face 25, is a skull, #27); none of the 32 Ghost testGANs is a closest piece for any face; and only six faces (1, 2, 7, 17, 20 and 24) have a clear nearest look-alike. The closest pieces are look-alikes, not identifications. Nothing was dropped, and the three best pieces per face are in the file. PW44 stays open: the faces are not Act 1 pieces, so the bitGANg's meaning is not settled by this match. - **What ran since the cycle 002 record: 5,255,990 phrases, every job complete, plant seen, standard path, no match.** The exhaustive ex/so pass finished: files 8 to 25 hold 1,796,119 phrases, and the whole pass is 2,496,119 phrases in 25 files (25 fewer than the 2,496,144 sets, because the build skipped sets already run). The cycle 002 leads' Tier A finished: lead 1, 25,428; lead 2, 24,570; lead 3, 149,700; lead 4, 62,131; lead 5, 61,355; lead 6, 997,595 (1,320,779 in all). The six held cycle 001 phrases ran as `c001_grids_eye_5_by_D4b`. Cycle 003's work by family, finished by 08:15 on 9 October (2,139,086 phrases): the rest is listed by pathway in the next point. - **Cycle 003's results by pathway (phrases finished by 9 October, 08:15; the pending file has one card each).** PW03, five drops open the sixth, 40,978. PW04, each drop holds a portion, 36,193, and the rebuilt word-list run across the drops, 36,300, plus its Tier B first files, 60,000. The Grader's other send-backs, 27,956, and the true palindromes on the Review, 78,467. PW41, the chapter titles, 15,663 (named forms and depth pass) and 120,000 of the cross tier. PW18, portal frames, 16,143, and powers of three, 5,221. The Points flip with its wrap stated (PW08), 2,121. PW23, the block code, 65,366. The Polybius reading, 162,712. PW31, each bitGAN one bit, 13,067. PW19, eyes in the background, 205,242. PW38, opposites, 5,564. PW46, the 813 mirror, 88,740 in Tier A and 357,420 in the first Tier B files. PW47, 181,403. PW48, 2,692. PW49, 23,094 and 83,774. The continued counts: the nine Ghost pieces #301 to #309, 2,102, and the eleven #301 to #311, 3,301. PW22, behind the binary, 4,844. The out-of-step eleven with 'random' last, 5,388. The Excavator's leads: the blocks read into the drops, 328,088 (115,064 Tier B), and the swap pairs, the colour-flash fourteen, the three exact elevens and the speed-up pools, 52,183. Every one: this exact combination did not open the wallet; its parts stay available. - **How the four batches were built and graded, with the rotation rule.** `asm_drops` (Fable 5.1) built PW03 and PW04: 540,052 phrases in 36 files; the Grader (Opus 5.5) rebuilt 12,025 of 14,400 sampled tasks with its own code, none failed, approved and queued 272,917 phrases in 25 files, and sent back hyp 7 in full (207,095 phrases; its Review share used #471 'someone', which is in list order, where the record names #458 'soup', #463 'sorry' and #494 'sort') and hyp 4's palindrome pools (60,040; the robots' six-bit palindromes are 0, 12, 18, 30, 33, 45, 51 and 63). `asm_empty` (Opus 5.5) built the empty pathways that needed no new parts: 1,278,859 phrases in 81 files; the Grader (Fable 5.1) rebuilt 4,553 phrases, found none already tested, and approved all 81 at odd priorities 501 to 681. `asm_mirror` (Fable 5.1) built invG's pathways, the send-backs and the Excavator's five leads: 3,489,497 phrases in 183 files (747,946 in Tier A); the Grader (Opus 5.5), who also reviewed build_003d, rebuilt 2,100 labels with no mismatch left, approved all 183 (queued as 191 files at even priorities 502 to 662) and dropped, rightly, the 1.8 million phrases of the continued-count key laid on whole drops as unguided brute force; all six send-backs of the drops grading came back and were queued. `asm_eye` (Opus 5.5) built from the 9 October eye answers: 285,808 phrases in 17 files (PW22 behind the binary; the GAN under the costume, read from the spooky pieces' Type and SubType; the continued-count eleven); the Grader (Fable 5.1) rebuilt 254 labels, approved all 17 and noted one weak point: the collector's own words call #141 and #129 sitting "behind the background" (not the binary), so stricter readings give an eleven and a ten, which are requested as a small labelled job. In every pair the grading model differs from the composing model. - **New per-piece facts** (anomaly ledger, 105 records on 82 pieces, source `system_2026-10-09`, each recomputed by script from `primary_facts.py` and the ArDrive records before writing). The mirror census: ten pairs of the 813 mirror (#512 on #301, #450 on #363, #473 on #340, #474 on #339, #497 on #316, and #501 to #505 on #312 to #308), 20 records, and the eleven Named Review pieces on exactly eleven Ghost pieces, 22 records. The continued-count eleven, Review #470 to #480 on Ghost #301 to #311, 22 records. #467, #468 and #469: the file's last-modified time on disk is 1 January 2022 (09:17:44 to 09:17:46 UTC) while the Photoshop save time is 23 and 24 October 2021 (99,930, 99,675 and 99,611 minutes earlier), where it matches within seconds for the other 209 of the 212 Ghost and Review files. Seventeen Ghost pieces were uploaded to Arweave twice (#307, #308, #309, #311, #312, #313, #315, #316, #317, #329, #336, #337, #338, #340, #341, #342, #344); an older note of nineteen including #403 and #440 does not match the records, which hold one copy each. The collector's eye answers of 9 October as his observations: five behind, one cannot tell, nine in front, and six portal-frame pieces with no target named, 21 records. One new tag, `behind_binary_by_eye`, the five Review pieces he marked behind (534 tags and 9,626 assignments on the local build, 533 and 9,621 before). - **Totals, counted by script on 9 October 2026 at 08:21.** Run since the cycle 002 record (6 October): 5,255,990 phrases in 231 job files, all with the plant seen and no match, and a further 60,000 phrases (three of six files of the costume batch, the GAN under the spooky costume) finished after this record's list was frozen and will be in part 2. Waiting in `system/queue/`: 4,833,012 phrases in 258 files, of which Tier C holds 23.8%, the cycle 003 mirror and excavation files 42.4%, the empty pathways 13.8%, the cycle 002 Tier B 11.5%, the eye files 4.5% and the two drop pathways 4.1%. The ledger's phrase table, opened read only, held 10,128,920 rows from 739 valid jobs (7,357,775 under the standard path label and 2,771,145 under an older path label, as `ledger.py stats` prints them). Nothing has opened the wallet. Site cards for the finished jobs (54 entries) are in `site_sync_pending.json` for the lead to sync. Details: `system/cycles/003/site_record.md`. - Next D-number: D126. ## D126 — Cycle 003 of the solving system, part 2: the cross and Tier B files, the PW19 cross, cycle 002's Tier B and Tier C, WordNet's five pairs, the site bridge, compression, two engine fixes, and 13,803,119 more phrases run (no match); the queue drained on 11 October (2026-10-11) Written by the Recorder of cycle 003 (part 2). Earlier entries are not rewritten. Sources: `system/cycles/003/assembly_asm_drops.md`, `assembly_asm_empty.md`, `assembly_asm_mirror.md`, `assembly_asm_eye.md`, `assembly_asm_wn.md`, `assembly_asm_pw19cross.md` and their gradings, `system/cycles/002/assembly_asm_leads.md`, `grading_leads.md`, `grading_tierC.md`, `system/cycles/003/build_site_bridge.md`, `build_build_003e.md`, `system/ARCHITECTURE.md` (section "9 October 2026: compressed job files"), `system/queue.log`, the finished jobs in `system/done/` and `python3 ledger.py stats`. The 34 site cards for these jobs are in `site_sync_pending.json` (one per pathway family or lead; the PW19 cross is one card for its 378 files); the lead runs `site_sync.sh`. - **What the collector said on 9 October, besides his eye answers (those are in D125).** About the PW19 cross, which the Grader of the empty pathways had put to him as a question (about 6.4 million phrases, about 19 hours): "run it". It was built as 7,543,756 phrases in 378 files (priority 750; pools of 13, 15 and 18 pieces from his notes on eyes and heads behind the main face, cut by 146 forms, every reader other than the essence crossed with every transform and every distinct order) and it ran. About the disk: "Let's compress and if we need to figure out a way to track what's been searched without necessarily storing it all." About the dictionary, "You can download WordNet." (D125 has his first question, what WordNet was); the result is two points below. About the site, in a question: "Are you actually updating the website? What's with the stack of untested hypotheses?" (the next point). - **The site bridge (Builder, 9 October; `cycles/003/build_site_bridge.md`).** Why the site did not show the system's work: the site works out its "Where we have looked" section and its tested and untested marks from its `runs` table. The system's finished jobs had reached the site only as theory cards whose run rows said "see card" and named no selection, word rule or order, so the site could not count them. What was built: `system/site_bridge.py` reads every finished job, parses each task's label, works out which pieces it used (from the registry when the pool is known, from a list or window in the label, or by turning the words back into pieces when the word rule is the plain essence), matches them to the site's own selections, word rules and orders by hand maps, and writes run rows. `site_sync.py` gained a `bridge()` step after `cards()`. A pool with no site selection becomes a new selection `SEL-SYS-...` when it was used in 10,000 phrases or more. For the collector's own example (the first and last piece of each drop), the bridge writes a run row for `SEL-DROP-ENDS-12` with 1,968 phrases. The dry run of 9 October counted 465 jobs, 9,128,471 phrases, 932 run rows, 22 existing site selections matched, 62 new selections, 17 new word rules and 11 new orders, and 527 pools that could not be resolved (413,714 phrases, which still get rows without a selection). A second fault was fixed: `coverage()` had reported "11177 new sets" on every run because it sorted one side's words and not the other's; both sides are now sorted, and the system's finished jobs are read one at a time (a dry run counted 5,692,132 new sets, about 402 MB, which is a large first upload against the data folder's 103 MB; `--no-system-sets` skips it). The site page `index.html` now also accepts `EC-SYS-` and `OC-SYS-` ids, which goes live with the next `update_all.sh`. Open limits the Builder listed: the plain text of the 17 rules now marked tested still says "Tested so far: no" or "partial" (a dated correction is a separate edit), and the coverage files are not tracked in the site's git folder (a careless `git add .` there would add 400 MB). `site_bridge_done.json` now lists 469 job names, which means the bridge has been run for real at least once since the dry run; this Recorder did not run or check the site. - **Compression and `system/jobfiles.py` (9 October).** Finished files in `done/` and the cycle folders' `hyp_*.json` are gzip-compressed in place; nothing is deleted and `gunzip` restores any file. `queue/` stays plain for the runner. `jobfiles.load(path)`, `jobfiles.glob_jobs('done/*')` and `jobfiles.stem(path)` read either form, and `precheck.py`, `leads_engine.load_seen`, `depth_pass.py` and the label census test use them. On 11 October `done/` holds 1,083 files, all `.json.gz`, about 408 MB. The ledger is the record of what was searched, so a lost job file would not lose that record: the ledger's phrase table held 23,892,039 rows from 1,448 valid jobs when read (21,120,894 under the standard path label and 2,771,145 under an older path label, as `ledger.py stats` prints them; 24 jobs invalid, 358 planted tasks; source `system/queue` 21,119,275). - **Two engine fixes of 9 October.** (1) A worker once read a job's plant file while it was empty and crashed the whole run with a JSON error; `engine.py` now retries the read up to 20 times, a quarter of a second apart, and stops with a clear message if the file is still unreadable. It also handles a plant file left empty by a kill in mid-write: it removes it and starts clean if no unit of that job is done, and stops for the lead to check if some are. The queue log shows the crash once, on 9 October after 22:56, in `c002_asm_leads_2_tierB_p07`; the job resumed from its saved units (39 of 50 done) and finished with the plant seen. (2) `run_queue.py` now refuses a job whose name has already finished (a `state/.result.json` exists): the file moves to `queue_refused/` with a log line, instead of being skipped without notice. The log holds no refusals, and there is no `queue_refused/` folder, so no job was refused. - **The WordNet result (`cycles/003/build_build_003e.md`, `assembly_asm_wn.md`).** The Builder froze the dictionary's direct antonym links between seed words in `vendor/wordnet_antonyms_2048.json`: 200 seed words have a direct antonym among the seed words, 86 have one by a synonym step, 263 either. Forty-five pieces carry a word with a direct antonym (skull 11, robo 5, spooky 1, Solstice 8, Ghost 12, Review 8) and 57 with a direct or one-step one. Only five pairs have both words on pieces: #5 above and #245 below (skull and Solstice); #78 answer and #367 question (skull and Ghost); #363 true and #473 false, #371 penalty and #496 reward, #382 multiply and #449 divide (all three between the Ghost drop and the Review). Compared with the hand list written for the card: the outside list confirms three of its ten Ghost and Review pairs, lacks sixteen of its twenty named pairs, and adds two pairs the card did not have. The collector wrote on 7 August that the true and false essences are "intentional", which stands beside one of the five. What ran: the five pairs and the three Ghost and Review pairs (10,721 phrases), the two WordNet pools cut by every reducer and read by the essence and the two opposite readers (652,709, Tier B), and the same cuts read by every other reader (442,632, Tier C); all of it finished with the plant seen and no match. These exact combinations did not open the wallet; their parts stay available. The five pairs were added to the anomaly ledger as scripted facts (source `system_2026-10-11`, 10 records). - **The bitGANg 1001 frame match.** Already recorded in full in D125 (no Act 1 piece among the faces, no exact copy, six faces with a clear nearest look-alike); nothing was added since. - **What ran since D125: 13,803,119 phrases in 711 job files, every one finished complete with the plant seen, on the standard path, with no match** (D125 counted 5,255,990 phrases in 231 files up to 08:15 on 9 October, so the whole run since the cycle 002 record is 19,059,109 phrases in 942 files). This count includes all six files of the costume batch (the first three, 60,000 phrases, were named in D125 as finished but not carded). By pathway: PW19 7,543,756 (378 files, 54.7%); PW38 1,204,385 (62 files, 8.7%); PW04 1,095,983 (58 files, 7.9%); PW08 967,475 (50 files, 7.0%); PW18 816,800 (44 files, 5.9%); PW03 524,579 (31 files, 3.8%); PW06 480,492 (25 files, 3.5%); PW46 451,227 (24 files, 3.3%); PW05 257,050 (14 files, 1.9%); PW41 233,639 (12 files, 1.7%); PW49 170,783 (9 files, 1.2%); PW22 56,950 (4 files, 0.4%). By family: the PW19 cross 7,543,756; cycle 002's Tier C 1,150,350 (lead 1 721,661, lead 2 47,093, lead 5 381,596) and Tier B 554,667 (lead 1 245,814, lead 2 209,957, lead 5 98,896); the WordNet batch (pairs, strict PW22, eight spookies, nine portal pieces, WordNet pools) 1,366,351; the mirror batch's Tier B piles 2,046,826; the empty pathways' cross tier 667,760 (PW41 233,639, portal frames 220,356, powers of three 115,442, opposites 98,323); the drops batch's Tier B 195,746; the eye batch 277,663. Every one: this exact combination did not open the wallet; its parts stay available. The queue drained on 11 October at 07:06 (the last file was `hyp_asm_leads_5_tierC_p20`); `system/queue/` is empty. - **Pathways with no phrases yet, as D125 listed them:** PW15, PW16, PW17, PW20, PW21 and PW24. No cycle 002 or cycle 003 job file carries one of those pathway names. - Next D-number: D127. ## D127 — Cycles 004 and 005 of the solving system: the Reframer's picks run, animation events, the Review's working groups, the collector's words of 22 to 25 October, the four step judges and their maps, the brainstorm plan, the workshop and designer censuses, the cycle 005 mechanisms, the pinned-order runs, and 16,323,595 more phrases run (no match) (2026-10-25) Written by the Recorder of cycle 005. Earlier entries are not rewritten. Sources: `system/cycles/004/assembly_asm_r1.md` to `assembly_asm_r5.md` and `grading_r1.md` to `grading_r5.md`, `build_asm_r4_rerun.md`, `cycles/005/assembly_asm_c5.md` to `assembly_asm_c5e.md` and `grading_c5.md` to `grading_c5c.md`, `STRATEGY_BRIEF.md`, `BRAINSTORM_PLAN.md`, `SYNTHESIS.md`, `WORKSHOP.md`, `CENSUS_DESIGNER.md`, `system/CANON.md`, `system/roles/judges.md`, `system/maps/*_BREADTH.md`, `system/queue.log`. Every miss below means only that the exact combination did not open the wallet; its parts stay available. - **The collector's words of 22 to 25 October, in his words (`CANON.md`).** On 22 October, about the 813 mirror pairs: #473 and #340 share only the space-invader icon ("sort of looks like it has an icon on the shirt, but it never fully materializes"); for #474 and #339, #512 and #301, and #450 and #363 he found "absolutely nothing" linking each pair, so the 813 mirror is not a visible rule by his eye. On 23 October, the 42 difference loops (each frame minus the one before) showed "I see nothing". He explained how the pieces are made: a GAN interprets static images, and the motion of the GIF is "the literal interpolation between frames", so a mark hidden in the change between frames is "a highly unlikely strategy"; only frames Pindar edited by hand (the portal frames, the Fibonacci-timed stutter on the prime skulls, the inverted frames on #231, #256, #274 and #479) carry anything planted in the motion. Same day, his correction: Pindar "controlled the training data to produce some of these visual effects embedded in the motion. They stand out as something intentional in the training data that drove the creation of the GIFs." So a visible motion effect can select a piece; it does not carry drawn digits or letters. On 24 October, his design answers: he would "repurpose many things from past drops" and "require multiple steps", where something from a drop's own puzzle or theme identifies a glitch and another step finds the word, and possibly "you needed the full candidate set to even uncover the transformation step". To find the glitches: "take the theme of the drop and see what doesn't fit", with four frames held at once (each drop alone; Ghost and Review as a pair; the pre-Review collection and the Review as key and lock; the whole collection). On errors: Pindar "made a lot of purposeful errors and a lot of careless unintentional errors. It's hard to know which is which and he could always coopt the unintentional ones." On the word: he leans against a word in the art; Pindar's later First Sparks puzzle "hid the words in the description"; candidates are the essence, another word in the name or description, a pointed piece's field, or a mathematical relationship among the glitches. On the order: "we always have to brute force all possibilities", and "the mint order on the contract is not the same as the ordering of the collection. Sometimes they are different and sometimes there are gaps as other things were minted in between the drops." On 25 October he corrected two things: by "name" he means the piece's TITLE (every piece has one), not the Name trait that only nineteen pieces carry; and the descriptions "are not the same in every drop and on every piece, even inside a drop", which led to his clue about a different spreadsheet (PW35), so the differences are to be dug into, not discounted. - **His standing rulings on how the system runs.** "split the pathways" (8 October; the queue alternates between the ex/so family and the pathways with no phrases). Tier C as filler, 24 October: "Let's make it so that you run tier C whenever nothing else is running. But I always want you working on new strategies to push this forward." The lead's reading: bulk piles are filler at priority 900 and above, new-strategy jobs take 500 to 699 and always run first. Always a strategy agent: "always keep an agent running reviewing results, thinking through new things while my CPU churns out calculations. Just keep it metered to my allowed Fable/Daily usage" (`cycles/README.md`). The two-tier spend, agreed the same day: the continuous loop is the cheap feed (Sonnet censuses, builders, job generation, one light grader per batch), and the heavy review (the Reframer and the full four-judge panel, Fable and Opus) runs only when a named-form batch finishes or the queue drains; if the budget tightens, the light loop slows first. On 25 October he approved the judges' proposal in two words, "Do the proposal", which started the pinned-order runs below. - **The four step judges and their maps (24 October; `roles/judges.md`, `maps/*_BREADTH.md`).** His words: "judges who can look at each step of the puzzle process, identifying the glitches, number one. Identifying the words, number two, and identifying the order, number three, and have a sense of the breadth of possibility inside each one". A fourth judge looks at how the three steps join. Each judge keeps a living map; the mechanical half is generated by `python3 system/step_maps.py`. Judge 1 (selection, Fable 5.1) counts 4,434 pool ids over 2,110 job files and 40.6 million phrases, with 2,095 registry pools never used, and groups the selection rules into families: a planted word run and its breaks; a count that fits eleven; a drop's theme and its misfits; a key from another drop laid as places; the collector's eye; and others. Judge 2 (words) counts the typed essence at 7 readers and 10.9 million phrases, another typed field read as text at 26 readers and 491,000 phrases (the Name and description are thin), a number read as a word number at 98 readers and 7.5 million phrases, the word that should have been there at 10 readers and 10,648 phrases, and a pointed piece's field at 35 readers and 1.5 million phrases (none returns the target's Name or description). Judge 3 (order, Opus 5.5) lists 108 order ids and 15 shape ids. Judge 4 (mechanics, Fable 5.1) regroups what has been joined with what: for example, the planted runs met the essence reader at 2.79 million phrases and had never met Pindar's own text as an order. Every verdict feeds the Grader and the Reframer. - **The brainstorm plan and its steps (lead, 23 to 25 October).** The strategy brief of 23 October (Opus 5.5) said about 30 million tested phrases mostly read the typed essence of eleven pieces in the Ghost drop and the Review, and proposed four strategies: A, a printed number that turns into the essence of a different Act 1 piece (the SuperRare closure); B, glitches that point at other pieces (the portal walk); C, eleven kinds of glitch rather than eleven pieces; D, words from Pindar's own sentences with the order from the pieces. It also proposed a difference loop for each event frame, and holds (animation-event Tier C, new mirror axes, the 'so' group Tier B in every order). The collector ran the loops ("I see nothing") and chose to keep the Tier C as filler. The plan (`BRAINSTORM_PLAN.md`): Step 0 the assumption ledger (the lead; `ASSUMPTION_LEDGER.md`); Step 1 three fresh-eyes solvers given only the public pack, who converged on elevens the ledger already held (the 'so' block and 'ex' column, the Ghost fixed points #307 #322 #348 #368 #372 #380 #389 #404 #407 #429 #444, the error cells, the Ultra and Silver elevens); Step 2 Pindar's workshop; Step 3 the collector as designer (his answers above); Step 4 the synthesis of five mechanisms; Step 5 build small, named forms only, with depth only on exact elevens and nothing over 100,000 phrases without his word. - **The workshop facts (`WORKSHOP.md`, Opus 5.5, 24 October).** Skulls were saved 20 to 27 August 2021, robos 24 to 27 September, spookies 15 October, Solstice 9 to 21 December. #467, #468 and #469 were made in October 2021 (saved 23 and 24 October, 81-frame testGAN files; every other Review piece has 55 frames and is from January). On 1 January 2022 one copy wrote all 128 skull files, all 64 robo files and #467 to #469 in eight seconds (their save times did not change), the five early zips were uploaded, and the sixteen spooky files were written one at a time in the order 195, 197, 193, 205, 199, 207, 208, 196, 200, 201, 202, 203, 206, 198, 204, 194. The Review was built group by group from 6 to 30 January, so the Review numbers were assigned after the files: the numbers follow the file-name working groups except at the seams (no reprise07; #465 and #466 are group 08; #470 to #472 and #477 are group 11; #478 is group 10). The last session, on 30 January, wrote #447 #446 #453 #454 #512 #460 #457 #461 #466 #465 #480 #479. On 31 January the sheet 'bitgans_act1.csv' was saved at 19:42, two minutes after the Ghost upload ended, and the 68 Review files were uploaded in four seconds. The 1 January table equals the 31 January sheet on every cell of #1 to #300 except the skull Drop cell, so every word in the first four drops was fixed by 1 January. The clue went live on 5 February at 12:12. - **The designer census (`CENSUS_DESIGNER.md`, four scripts, 24 October).** `TI` in `primary_facts.py` is the contract's own token index (the mint slot), read from `token_index.json`; it is not a rank among the 512. The 512 pieces sit on slots 66 to 662, so 85 slots inside that span belong to something else: 67 gaps (58 of one slot, five of two, two of three, one of four, one of seven; the sizes 1, 2, 3, 4 and 7 are the words abandon, ability, able, about and absorb) hold 85 tokens, of which 74 are unreadable and eleven are readable (eight reneGANs 0001 to 1000, 'Lexy 20,200', 'Sarah 15,416' and 'Frozen ghostGAN (90/144)'). Ranked by slot, 164 pieces differ from their number, but only eight are truly out of place (#24 #30 #310 #311 #359 #390 #445 #446); the other 156 shift by one because of them. The six backward neighbouring pairs are (#24,#22), (#30,#29), (#315,#310), (#444,#390), (#390,#359) and (#511,#445); they hold eleven different pieces. Descriptions: every piece has a full chain description with 7 to 15 seed words (the drop's standard text holds 9 to 11); 15 pieces differ from their drop's text (the same 15 as `DESC_ANOMALY`); 169 pieces carry exactly one extra seed word in title, name, chain name or file name, but only 33 distinct words. Pointers: only the drop counter and the 513 mirror give every one of 66 pool members a target with a usable word. The set's own mathematics: the 'so' block's steps are 3 5 5 8 8 5 3 7 5 3 6, so nine of its eleven differences are Fibonacci numbers, and the 'ex' column's steps are 12 with one 13. - **The mechanisms built.** Builder build_004a (`parts/cycle004a.py`): the read_down shape family (first eleven distinct, ten plus 'final', twelve less the last) and sixteen field readers. Builder build_004b (`parts/cycle004b.py`): 46 animation-event readers and the event pools. Builder build_005a: nine pointer rules with four readers each, and 18 keys computed from the chosen set (shift, flip and at_key_places on the set's own differences). Builder build_005b (`parts/cycle005b.py`): the Title readers, the edit and difference word readers, the pool `desc.differs` (15 pieces) and `read_at_place`. The engine kept its two fixes of 9 October and the patched build scripts of 11 October (nothing deleted). Part requests still open: a pointed edit-word reader, a pointed word-at-its-place reader, a `key_from=` setting and a pointed Title-word reader. - **Cycle 004 results (built 11 October; every batch graded by the other model).** asm_r1 (Opus 5.5, graded by Fable 5.1): WordNet opposites cut by drop (the Ghost twelve had never run), the owed small jobs and the continued counts, 1,433,838 phrases in 86 files. asm_r2 (Fable 5.1, graded by Opus 5.5): a field read down a drop with repeats dropped; the headline reads the Ghost titles in mint order, giving slender vivid frozen twin dry shy eye swarm test spin space; new pathway read_down_field (proposed PW50). asm_r3 (Opus 5.5, graded by Fable 5.1): the Review's own numbers as keys into the other five drops, and the 825 mirror between 'seed' and 'random' with the Fibonacci cuts of the window. asm_r4 (Fable 5.1, graded by Opus 5.5): animation events (new pathway animation_event); a disk shortage cut the first build short and the Grader found two jobs never built and 181 census parts missing, so a rerun rebuilt them as new files. asm_r5 (Opus 5.5, graded by Fable 5.1): the collector asked whether the 'so' words had been mapped to the Arweave groupings; the answer was no, so the 25 working groups became parts for the first time. Each of the twelve 'so' words sits in a different group; the clue word 'random' (#468) is in reprisetest, the same group as 'soon' (#484). - **Cycle 005 results (built 23 to 25 October).** asm_c5 (strategies C, D and B, 15,824 phrases), asm_c5b (the five mechanisms of the Synthesis on pathways c005_S1 to c005_S5, 44,597 phrases), asm_c5c (the Reframer's five multi-step recipes and four owed jobs, 278,627 phrases), asm_c5d (the collector's design answers laid on the verified exact sets; a pointed Name reaches too few pieces and a pointed description gives the drop's shared first word, so only 2,843 phrases could be built) and asm_c5e (the Title-word pool of 132 pieces, the 15 description differences and the January-sheet disagreement five as a key; 172,456 phrases in 13 files, queued, not finished at the cut). asm_c5 to asm_c5c were graded by Fable 5.1 (Opus 5.5 composed them); asm_c5d and asm_c5e were queued by the lead without a separate Grader, to be judged at the next heavy review under the two-tier rule. - **The pinned-order runs ("Do the proposal", 25 October).** `cycles/005/run_perms.py` makes, through `permute_orders.py`, every order with both ends pinned in mint order (9!, 362,880 orders each) for the lying-label eleven, the Ghost fixed points, the mint-inversion eleven and the 24 'ex' and 'so' elevens (each block minus one piece), then every order with the last word pinned (10!) for the lying-label eleven; the nine on-purpose chain glitches (D90) were added as nine plus 'final' (priority 627). Finished at the cut: the whole lying-label 9! run (19 files, 362,875 phrases) and 4 of 18 files of the Ghost fixed points (80,000 phrases). - **What ran since D126: 16,323,595 phrases in 885 job files, every one finished complete with the plant seen, on the standard path, with no match** (counted from `queue.log` between 07:07 on 11 October and 20:40 on 24 October, because `ledger.py stats` was locked by the live runner). By script: cycle 004 batches 15,538,829 phrases in 821 files; cycle 005 batches and the pinned-order runs 784,766 phrases in 64 files (this includes the cycle 004 read-down Tier C and the asm_r5 send-backs, which were built in cycle 005). Of these, the animation-event Tier C filler is 7,077,914 phrases in 356 files; the pinned-order runs are 442,875 phrases. Not run: two cycle 004 Tier B files (`780_c004_asm_r4_ev_fib_loose_tierB_p08` and `_p17`, 40,000 phrases) were refused by the engine as already finished and sit in `queue_refused/`; the Tier C filler of the fib_loose, flash and census piles was still waiting. The whole run since the cycle 002 record is now 19,059,109 + 16,323,595 = 35,382,704 phrases. - Next D-number: D128.