# Confirmed data-glitch collection (canonical registry)
Updated 2026-08-25. Standard: byte-level or structural evidence, verified, no interpretive leaps.

## Confirmed
1. **#14 Rainbow skullGAN (ess `accuse`)** — identity corruption. Byte-identical to #75's file
   (MD5 92ac390f7fb463e640cdc6f3a74186df, shared XMP). Art is flaming-styled → #14 wears #75's art;
   its true rainbow animation is public nowhere. Extends to OpenSea (same CDN content hash).
2. **#75 Flaming Glitch skullGAN (ess `announce`)** — the other half of the pair (its art is the shared one).
3. **#390 Frozen ghostGAN (ess `century`)** — double-mint: slots 519 (trait-less) + 583, same image.
4. **#479 Inverted glitchGAN (ess `song`)** — timing glitch: all frames 0ms except frame 27 held at 100ms.
5. **#480 slenderGAN (ess `fossil`)** — timing glitch: all frames 0ms except frame 24 held at 100ms.
   Same mechanism as #479; the pair defines the class.

## Candidates (not confirmed)
- #146 G1itch1ng roboGAN — held frame at index 0 (common encoder quirk; robo glitches are authored-in).
- #228 Kevin aboschGAN — unique 50ms speed (only file in 512); Artist tribute; five-way encoder signature.
- Re-saved-with-no-detectable-change residue: #203, #407, #454, #457, #460, #461.
- SubType-vs-art contradictions (visual worker, family-calibrated z; instrument validated by recovering #14 as
  a top-2 Rainbow outlier): strong = #179 (Rainbow label, colorless ghost art), #423 (Rainbow, monochrome blue),
  #231 (Frozen, warm + mutating), #511 (Inverted on white bg, ess `negative`), #509 (Dark Rainbow, Inverted on
  light bg); medium = #204, #150, #490, #478, #92, #491. #150↔#179 are mutual label-swap partners
  (misfit 6.65→1.97 if labels exchanged) — consistent with one robo-drop swap event.

## Cleared (former suspects)
- #474 (quiet-frame detector false positive), #5 (loop-wraparound artifact).
- Montage no-match six {143,155,187,223,294,357} — RESOLVED as wrong-frame artifact: full search of all 26,781
  frames (with mirror/negative transforms, 22 calibration controls) puts every tile's global best match at its
  OWN animation, rank 1 of 512, inside the true-match distance band. The tiles show own art at mid-morph frames
  our earlier candidate-frame sampling never included. No tile wears another GAN's art.

## GANg-vs-glitch mislabel pair (2026-08-26, the collector's eye + census)
- **#281 "Entangled glitchGAN" (ess `book`, solstice, mint 387):** exactly 8 pieces have GANg
  literally 'glitchGAN'; 7 visibly glitch, #281 is the SOLE non-glitcher (the collector confirmed by
  eye). A clean 1-of-8 singleton. (Note: `book` is one of the unendorsed community
  potato/elite/book/moon words — logged, not endorsed.) NOT re-saved.
- **#47 "Twin Glitch skullGAN" (ess `alarm`):** the mirror — NOT GANg-glitchGAN, yet DOES glitch
  (cut dialect), and silently re-saved. 
- The pair: opposite GANg-vs-behavior mismatches. Candidate matched-pair signposts.

## SPOOKY DROP INVESTIGATED (fresh deep-dive, 2026-08-26) — theories/spooky_investigation.md
Every channel provably INTACT except titles: 16/16 uniform (55 frames, 100ms, identical block
structure, consecutive BIP-39 essences glue->grape no skip, perfect octal numbering, HiddenType
now 16/16 with #199=E). So spooky's glitch lives in its TITLES (the medium the drop is about).
- HYPOTHESIS: spooky's glitch = a one-character CASE corruption of the displayed name.
  CARRIER = #205 "Flaming spookyGan (0o14)", ess `grace` (lowercase g; sheet+chain identical;
  also its costume descriptor Flaming==its SubType Flaming, a 2nd name-layer fault). Case-error
  class in all 513 = exactly {#175 skullGan, #205 spookyGan}.
- NEW double-space channel: 6 of 16 spooky titles carry a double space (0o1,0o11,0o12,0o13,
  0o16,0o17) vs 1 in the other 497 (#174). As an 11-bit read: `attack` fwd, `scare` rev
  (caveat: multiple-comparisons-prone). If spooky's glitch is a whole-drop bitstring, word may be `scare`.
- #200 Count DEMOTED: it visibly glitches (scanline tearing) so Type=Glitch is honest/authored,
  not the hidden defect. #203 confirmed silent re-save (23:24:50 Oct 15, dotted PS install, pre-mint).
- the collector's-eye flags: #208 (least-ghostly "Ghostly", dense red Flaming), #207 (2 eyes vs sheet Eyes=3).
- BEST SPOOKY CANDIDATE: #205, ess `grace`.

## FEELING-NAMES ELEVEN (the collector's idea, 2026-08-26) — essence swept negative
Pieces named for human emotions/states: the 4 Shy {338,340,355,357} + awkward #141, anonymous
#195, friendly #203, angry #204, hungry #259, gazing #290, staring #451 = exactly 11, spanning
5 drops. All essences BIP-39: awkward goddess gown grab bless boy yellow theory off seed daughter.
Notable: #357 Shy skullGAN essence = `seed` (the puzzle's own word); #203 Friendly / #204 Angry
Clown = adjacent opposite-emotion pair. Full stateable depth, 1758 orderings, plant recovered,
NO HIT. The feeling-WORDS register is mostly non-BIP39 (anonymous/friendly/angry/hungry/gazing/
staring/shy all fail) so not sweepable as words. Semantic-selection axis, novel but essence-dead.

## CONFIRMED GHOST-DROP GLITCH: #357 (anchor, 2026-08-26)
#357 "Shy skullGAN (57/144)", essence `seed`, ghost drop. Confirmed as a unique color-sequence
glitch: predominantly YELLOW where the ghost drop's rainbow gradient calls for green/cyan — a
step BACKWARD. OVER-DETERMINED across SIX independent channels: (1) the collector's eye (original +
composite recheck), (2) Claude's eye (composite concur), (3) both color metrics (full-picture
+figure), (4) montage no-match grading, (5) essence literally `seed` (the puzzle's own word),
(6) a "Shy" feeling-name skull. STATUS: fixed ANCHOR — treat as almost-certainly one of the
eleven; any candidate set excluding #357 is suspect; prefer constructions that naturally include
it. Ghost drop's glitch medium = COLOR (theme confirmed), but color is too noisy to SELECT a
clean eleven or to READ as words; it serves as a tiebreaker + this one anchor.
Companion anchor: #512 `final` (only Dim-11 piece, pinned last slot).

## TYPE-RULE-BREAKER CLASS (the collector's framing, 2026-08-26) — type_rule_breakers.json
A piece that FAILS TO DO what its type/name requires is glitching by omission — the rule-break
IS the glitch. Unifies three previously-separate findings: (1) behavior-omission (#281 glitchGAN
that doesn't glitch, 1-of-8; #47 the inverse), (2) SubType-vs-art / name-vs-reality (the signature
violators: #14,92,150,179,204,231,423,478,490,491,509,511), (3) misplaced/wrong-drop identity
(#175,187,191,206) + naga/twin breaks (#181,192). 20 pieces, ALL SIX DROPS. Tier-1 (visual /
theme-break) per the collector's priority. Note the robo trio #187/#191/#192 carry meta-essences
mirror / err0r / zero. Pool for Sudoku slot-filling; not itself the eleven (too many), but a
prime Tier-1 source. #281 confirmed here as a type-rule-breaker glitch (paired-class member).

## LATENT CLASSES from one-offs (the collector's method, 2026-08-26)
Directive: individual anomalies that seem like one-offs often form a hidden class. Apply this lens.
NEW CLASS found: NON-BIP-39 ESSENCES = exactly 3 in all 512: #191 `err0r`, #363 `TRUE`, #473 `FALSE`.
Computational/boolean non-words deliberately placed where BIP-39 essences belong; TRUE/FALSE = a
boolean pair, err0r the third. Each was an isolated oddity; together a clean 3-member class.
(Note: #151 `satoshi` IS valid BIP-39 despite the collector's note - a bitcoin reference, not a corruption.)
PAIR PATTERN: most singleton anomalies are actually 2-member pairs: day/night #209/#300, beige
#479/#480, scanline #489/#491, sine-wave-flash #130/#133, painted-numbers #165/#207. Pindar builds
in deliberate pairs. Standing worker directive: cluster one-offs by shared rule before treating any as unique.