# Lens 3 — Act 1 Review (#445–512) as the answer key Idea agent report, 21 September 2026. Ideas only. No wallet was derived, the sweep harness was not run, nothing on the network was touched, and no existing file was edited. The only arithmetic I did beyond reading files was the BIP-39 checksum (a hash of the word numbers), which is the same "checksum arithmetic only" step used in D75. Every piece fact below was read from primary_facts.py, lore_catalog_runs/gif_facts.json, lore_catalog_runs/deep_review_1/data.json or tags.sql. Every word set was looked up in theories/_registry.json. ## What I found before ranking anything (four new facts, all checked against the files) 1. **Exactly twelve Review files break the Review's file standard.** The standard is 55 frames with every frame shown for 100 ms. 56 of the 68 Review files obey it. The twelve that do not are #445 and #446 (88 frames), #467, #468, #469 (81 frames), #479 and #480 (one held frame, the rest at 0 ms), #493–496 (89 frames) and #512 (63 frames). One plain test, no threshold, and it returns twelve. The Spooky drop and the Ghost drop contain no file that breaks the standard; the Robo drop has 17 and the Solstice drop has 4. Earlier work listed the ten odd frame counts and stopped at ten (theories/review_groupings.md section 5, card 199). Nobody added the two held-frame files under the same test. 2. **Those twelve essences, in mint order, pass the BIP-39 checksum as a complete twelve-word phrase, and it ends on 'final'.** Mint order is #467, #468, #469, #479, #480, #493, #494, #495, #496, #445, #446, #512. The words are: give, random, calm, song, fossil, bronze, sort, erase, reward, collect, rose, final. A fixed twelve passes the checksum by luck one time in sixteen. 3. **In that same phrase, 'song' sits in place 4 and 'sort' sits in place 7, and those are exactly their places inside the solution block.** The solution block in wordlist order is solution 1, solve 2, someone 3, song 4, soon 5, sorry 6, sort 7, soul 8, sound 9, soup 10, source 11, south 12. #479 'song' and #494 'sort' are the only two pieces that are both in the solution block and in the file-breaking twelve. The chance that two named words land on two named places out of twelve is 1 in 132. 4. **#479 plays the Solstice negative-frame glitch backwards.** #231, #256 and #274 share one list of eleven colour-negative frames: 37, 38, 43, 44, 47, 48, 50, 51, 52, 53, 54. #479 has thirteen: 1, 2, 3, 4, 5, 7, 8, 11, 12, 17, 18, 26, 27. Take each of the trio's frames away from 55 and you get 18, 17, 12, 11, 8, 7, 5, 4, 3, 2, 1. That is #479's list exactly, with two frames added (26 and 27, and 27 is #479's held frame). The Review piece is the earlier glitch run in reverse, to the frame. The ledger records all four lists but not this mirror. Honest note on the checksum passes. I tried about thirty orders of the twelve in fact 1 (piece order, mint order, save order, each both ways; the twelve rotations of mint order and of piece order; four versions with 'random' moved to the end). Four passed; chance predicts about two. So a checksum pass is not evidence by itself. Its only value is that it names an exact phrase that costs one wallet test. Fact 3 is a separate coincidence and is the one I would weigh. --- ## Idea 1 (best) — The twelve Review files that break the file standard, in mint order **Side:** all DATA (file bytes and mint slots). **(a) Selection.** "The glitches are the Review files that do not run the way every other Review file runs: 55 frames, 100 ms each." That gives #445, #446, #467, #468, #469, #479, #480, #493, #494, #495, #496, #512. **(b) Word.** "Each piece gives its essence as printed on my sheet." **(c) Order.** "Read them in the order I minted them." That is the same order rule as Pindar's two solved puzzles. **(d) Shape.** Two readings, both cheap. Reading A: the full twelve, with nothing added: give random calm song fossil bronze sort erase reward collect rose final. Reading B (closest to the clue): the first eleven in mint order, give … rose, followed by a random twelfth word; 'final' is one of the legal twelfth words, so reading A is contained in reading B. **What in the data teaches it.** The Robo drop teaches that a broken file marks a special piece: its 17 file-breakers include the six palindrome-named robots (#129, #147, #159, #162, #180, #192) and #146 'G1itch1ng roboGAN', which has 63 frames and a held first frame. The Solstice drop's four file-breakers are exactly its four named artist tributes (#228, #246, #264, #282). Then two whole drops (Spooky, Ghost) have none, and the Review has twelve. #512 'Quantum glitchGAN' has 63 frames; the only other 63-frame file in the collection is #146 'G1itch1ng roboGAN'. The three October 2021 tests begin the mint-order phrase with "give random", and #445, #446 and #512, which Pindar chose to mint last, end it with "collect rose final". **What it predicts that we can check without a wallet test.** - The 4th and 7th places hold the 4th and 7th solution-block words (already true; fact 3). - The phrase passes the checksum with no twelfth word chosen (already true; fact 2). - If the solution block is a place ruler, then some relation should tie the other ten pieces to the so-word of the same place: solution–give, solve–random, someone–calm, soon–fossil, sorry–bronze, soul–erase, sound–reward, soup–collect, source–rose, south–final. 'south' (the bottom) against 'final' (the end) is the only one I can see. The collector's eye on these ten pairs is the check. - The rotation of the same ring that ends on the word 'random' also passes the checksum: calm song fossil bronze sort erase reward collect rose final give random. That reads the clue literally: eleven glitches, ending with the word 'random'. A GIF is a loop, so a ring with a marked start is not a strange idea here, but I hold this one loosely. **Already tested?** No. The registry has none of the twelve possible elevens from this twelve (all twelve looked up). 'universe' + the ten odd-frame pieces was tested (hyp79, card 199). The eleven 'wrong speed' pieces are a different set. D87 and review_groupings.md record the ten and stop. **How many phrases.** Reading A is one phrase. With the three other checksum-valid orders it is four phrases. Reading B in full (leave each piece out in turn, 12 plain orders, 128 twelfth words) is about 150 elevens. **Why Pindar would do it this way.** He builds each GIF by hand in Photoshop, so frame count and frame timing are his own deliberate choices, and they are the most literal meaning of "glitched file". The rule needs no judgement: anyone can open the 68 files and count. It also explains "only solvable after some more drops": the glitched files did not exist until the Review, and the essences were not public until the sheet was. **Weakest point.** The clue says eleven glitches plus a random word, and this is twelve glitches ending on 'final'. Reading B repairs the count but then #512 is not one of the eleven. The 89-frame and 88-frame files may simply be "high definition" work habits, not puzzle choices. --- ## Idea 2 — The eleven working groups that have no 'so' word, closed by 'final' **Side:** all DATA (Pindar's filenames and sheet). **(a) Selection.** "Twelve of my working groups carry a word from the solution block, so they are solved. The groups with no such word are the glitches." Fold the one-off plural 'repriselines' (#486) into 'repriseline' and there are 24 groups: 12 with a so-word, 11 without, and #512's own group. The eleven are: colborn #445, roses #446, brinkman #447, stella #448, reprise01 (#451, #452), reprise03 (#455, #456), reprise05 (#459–461), reprise08 (#465, #466), reprise09 (#473–476), reprise10 (#478), skullall (#489–492). **(b) Word.** "Each unsolved group gives the essence of its first file." That gives collect, rose, cloud, legend, daughter, trade, small, tragic, false, recycle, draw. (#473's essence is printed FALSE; 'false' is a seed word.) **(c) Order.** "Piece order", which is also Pindar's group order. **(d) Shape.** Eleven words, then the twelfth. With 'final' (the twelfth group) as the twelfth word, the phrase passes the checksum in piece order: collect rose cloud legend daughter trade small tragic false recycle draw final. I checked four versions for this (first file or last file of each group, piece order or mint order); one passed, which is about what chance gives, so treat it as a named phrase and not as evidence. **What in the data teaches it.** D79 listed the groups without a so-word as thirteen and saw no pattern. The count becomes eleven plus 'final' once the plural filename is treated as the slip it looks like. Several of the first files are the pieces this week's same-name work already flagged: #455 (the Review piece in the Flaming skullGAN family), #473 FALSE (a totally new kind, per the collector), #478 (joins the Rainbow bitGAN family), #489 (shares the name of #411 'key' and #450 'solution'). #461 'buyer' (odd five ways) is in group 05, so the in-group rule is the soft part. **What it predicts without a wallet test.** If the split into solved and unsolved groups is meant, the unsolved groups should look different to the collector's eye: for example, the unsolved groups may be the ones with no visible frame glitch, or the ones whose pieces match an earlier family. One look at the eleven groups side by side answers this. **Already tested?** No. Neither the first-file eleven nor the last-file eleven is in the registry. D79 says nothing was run on working groups beyond card 189. **How many phrases.** One phrase for the stated rule. All one-per-group choices: 2 × 2 × 3 × 2 × 4 × 4 = 384 elevens, times two orders. **Why Pindar would do it this way.** He filed the Review in these folders himself, and he placed twelve consecutive 'so' words one per folder. Marking the solved folders is a way of pointing at the rest. **Weakest point.** "First file of the group" is my choice; the data does not force it. The four tributes count as four groups only because each has its own filename tag. --- ## Idea 3 — LESSON plus a selection: the eleven negative frames are a stencil, and the Review reads it from the front **Side:** all DATA (file bytes). **The lesson.** #231, #256 and #274 carry the same eleven colour-negative frames, packed at the end of the loop. #479, the Review's Inverted glitchGAN, carries the same eleven mirrored to the start of the loop (fact 4). So the Review does not copy an earlier glitch; it runs it backwards. This is the first measured case of "the glitch teaches, the Review applies". **(a) Selection.** "The eleven glitch frames tell you which Review pieces to take: count into the Review as #479 counts into its loop." Places 1, 2, 3, 4, 5, 7, 8, 11, 12, 17, 18 of the Review are #445, #446, #447, #448, #449, #451, #452, #455, #456, #461, #462. **(b) Word.** "Each gives its essence": collect, rose, cloud, legend, divide, daughter, what, trade, coffee, buyer, never. **(c) Order.** Place order (the order of the frames). **(d) Shape.** Eleven plus a random twelfth. **What it predicts without a wallet test.** The stencil skips places 6, 9 and 14, which are #450 'solution', #453 'solve' and #458 'soup': all three so-words inside the stencil's reach (places 1 to 18) fall in its seven holes, which happens by chance about one time in twenty-three. It also overlaps Idea 2 heavily: it takes all four tributes, both pieces of groups 01 and 03, and #461 'buyer'. A second check: the collector can look at whether other Review pieces with the standard frame glitch also run their earlier partner's glitch in reverse (#503, #512). **Already tested?** The Review reading is not in the registry. The same stencil laid on the Skull drop (#1, 2, 3, 4, 5, 7, 8, 11, 12, 17, 18) is already tested (reprise_sets.json). The unmirrored stencil on the Review (#481, 482, 487, 488, 491, 492, 494, 495, 496, 497, 498) is not tested. **How many phrases.** Four elevens (mirrored and unmirrored, counting from 1 or from 0), a few orders each: under 50. **Why Pindar would do it this way.** "11 glitches" may be meant literally: the glitch he drew three times in the Solstice drop has exactly eleven frames. A frame list used as a pick list is the same kind of move as his earlier "right number, wrong list" puzzle. **Weakest point.** Laying frame numbers on piece numbers is a leap the data does not make for us, and the stencil is only known from four files. --- ## Idea 4 — ORDER LESSON: the solution block is a place ruler (the word's rank is the slot) **Side:** all DATA. This is a lesson about ORDER, not a new selection. **The lesson as one sentence.** "A piece that belongs to a solved working group takes the place, 1 to 12, that its group's so-word holds in the wordlist: solution 1, solve 2, someone 3, song 4, soon 5, sorry 6, sort 7, soul 8, sound 9, soup 10, source 11, south 12." **What teaches it.** The Skull drop teaches that a word's number is a piece's place (essence of skull n = word n). The solution block is the one place in the Review where essences run on consecutive numbers, 1655 to 1666, so it is the one place in the Review where that Skull-drop lesson can be applied. Fact 3 is the support: in Idea 1's phrase 'song' is 4th and 'sort' is 7th without anyone arranging it. **How it changes a test.** hyp74 used the block as a shuffle applied to seventeen elevens that had nothing to do with the Review groups. The ruler reading is narrower: it only orders words that come one from each solved group. The clean selection is "one piece from each of the twelve solved groups, chosen by one rule". Rules worth stating: the piece beside the so-piece; the group's first file; the group's piece that repeats an earlier name. I checked the two "piece beside it" twelves in ruler order; neither passes the checksum as a fixed twelve, so they would need the eleven-plus-random shape (leaving out the 'solution' group as the heading). **Already tested?** The ruler as an order for one-per-solved-group selections: no. The right-hand and left-hand neighbour elevens are not in the registry. **How many phrases.** About a dozen elevens per in-group rule. **Why Pindar would do it this way.** It gives 'sort' a job: the block itself is out of wordlist order in exactly three places (soup, sorry, sort), and sorting it is what turns it into a ruler. **Weakest point.** It rests on one coincidence (fact 3), and the in-group rule is open. --- ## Idea 5 — The Review pieces that repeat exactly one earlier piece's name, read on the Review side **Side:** all DATA (names on the sheet). **(a) Selection.** "The glitches are the ten names that one earlier piece and the Review share." Seven exact pairs give #463, #505, #484, #483, #485, #470, #481. Three names have two Review pieces each: #479 or #503, #472 or #486, #450 or #489. **(b) Word.** "Each gives its own Review essence": sorry, elder, soon, sphere, deposit, number, price, then song or broken, snake or remember, solution or draw. **(c) Order.** Piece order or mint order (the same thing inside the Review apart from the tributes). **(d) Shape.** Ten words, then 'final', then a random twelfth. **What teaches it.** Pair P01 (#14 shows #75's file) teaches that a repeated thing is a signal. The collector's working idea is to swap in the earlier word, and that side has been run (flame, beef, bike, bitter, exact, derive, key, then, mystery, circle; hyp74–76, hyp78, hyp79). The Review side of the same ten names has not been run. **What it predicts without a wallet test.** If the Review side is the answer side, the three two-piece names should each have one piece that stands out to the collector as the "real" reprise. For Inverted glitchGAN, fact 4 already picks #479 over #503. **Already tested?** No: all eight Review-side choices with 'final' are absent from the registry. **How many phrases.** 8 choices × a few orders × 2 shapes: under 100. **Why Pindar would do it this way.** A review that re-uses a name is the plainest kind of "review". **Weakest point.** Ten is reached only by counting names, not pieces, and nothing here is a glitch in the collector's strict sense. --- ## Idea 6 — LESSON (visual): the three early tests teach that eyes name a drop **Side:** all VISUAL. This needs the collector's eye before any test. **The lesson.** #467, #468 and #469 were made in October 2021 and wear spooky, skull and robo eyes: the three drops that existed. #467's eyes are those of #207, the piece with the painted '4'. The three tests have 81 frames, and '81' is the number painted on #165, a robo. So the tests may be pointing at the embedded-number pieces, and the question for the collector's eye is whether #469's robo eyes are #165's eyes and which skull #468's eyes come from. **(a) Selection.** "A Review piece that wears another drop's eyes is a glitch." Known so far: #447 (spooky eyes on an artist tribute), #506 (robo eyes on a bitGAN), #509 (skull eyes, robo mouth), #508 (eyes never the robo dot), and the three tests. **(b) Word.** Essence. **(c) Order.** Mint order. **(d) Shape.** Decided by the count the collector reaches: eleven, or nine or ten with the bookends. **What it predicts without a wallet test.** A census by eye of all 68 Review pieces lands on nine, ten or eleven wrong-eyed pieces. If it lands far away, drop the idea. **Already tested?** No such set is recorded. **How many phrases.** One eleven, 12 orders. **Why Pindar would do it this way.** The eyes are the one part of each GAN he places by hand. **Weakest point.** No count yet; it may be a family like 'anchored zero eyes' and nothing more. --- ## On the other lens items - **Missing reprise07.** Earlier work ties folder 07 to the twelve SuperRare bitGANs, and #445 was uploaded in the gap where 07 would sit. I add one reading only: #446 is titled 'SuperRare roseGAN', so the two tributes minted last (#445, #446) are the natural fill for the 07 slot. This matters for ORDER only: it gives a third Review order (tributes read at slot 07) to try on Ideas 1 and 2 beside piece order and mint order. - **Bookends.** In Idea 1 'final' closes the phrase with no bookend rule needed. #0 'universe' has no part in any idea above. - **'solution' → 'key'.** I found no new use beyond Idea 4, where 'solution' is place 1 of the ruler.