# SCION 52 · Gilliam Ephemera — reconstruct the method

PAE-02, the threshold, handmade. One agent working alone, with the SCION 52 prompt and `BRIEF.md`.

## What this is, and what it is not

A reconstruction of a **cut-out collage method** (photographic prints cut with scissors, hinged with pins, slid and
swapped on a lit table under a rostrum camera, shot a few frames per placement) applied to the threshold scene, built
only from public-domain prints in the kit and card I made.

- **No archive was consulted.** The research starting point in the prompt (a 1974 BBC demonstration on YouTube) was not
  reachable from here and I have not seen it, nor any Gilliam film or the SCION 01 film. Everything below about "the
  method" is **inference** from general, widely described cut-out practice (pinned joints, hinged jaws, substitution of a
  whole card for a turn, a mirrored print for the other side, held frames, airbrushed backgrounds), not a claim about what
  that demonstration shows. I claim no ephemera and no Gilliam image: none is in the film.
- **No paper was cut by hand and nothing was filmed on a real table.** The prompt's own hard proof (a physical paper
  gesture, filmed, then matched digitally) was replaced by the brief's scene, and it could not have been done here anyway.
  The tabletop is digital: pieces are raster cut-outs moved by rigid transforms only (see the protected conditions).

## The studio (smallest capable swarm, merged into one agent)

| role (prompt) | what I actually built or checked |
|---|---|
| broadcast source researcher | could not reach the broadcast: marked uninspected; no claim made from it |
| tabletop craft observer | a cut-out lexicon written as code (`build/table.py`, `build/puppets.py`): pin, slide, swivel, flop, substitute, hold, expose |
| cutout fabricator | `build/cut_pieces.py`: every piece cut from a named source frame by a scissor polygon (and its SAM matte), hand-tinted, paper core left at the edge; `proof/cutting_sheet.jpg` is the reconstruction drawing (pieces with their pins) |
| digital method translator | `build/xsheet.py` (exposure sheet), `build/render.py` (rostrum camera), `build/world.py` (set and lamps) |
| shot historian | `xsheet.json`: one row per exposure, every card's placement; `assets.json`: every piece's source print and rights |
| independent authenticity critic | **not me.** The builder cannot approve; an independent critic judges the film |

## Native formal language, as it exists in the project

- **Source prints**: `sources/` holds copies of the 17 frames used (dog, door, face, forearm, lamp, trees, cloud). Each
  piece in `assets.json` names its frame, archive URL, clip id and rights (all public domain), and its cut polygon in
  source pixels.
- **Physical materials**: prints are toned and hand-tinted (a three-stop dye ramp and an uneven wash), cut tight with the
  light paper core showing at the scissor edge, and cast a small shadow on what lies under them. Card pieces I made
  (crate, shoes, moon, black paper behind the jaw) are listed as drawn.
- **Cuts / pivots**: the dog is one print (home movie frame f006) cut into twelve pieces (round 3 repair): head, ear, torso, tail,
  and each of the four legs in two at the wrist or hock, with pins at neck, shoulders, hips and tail root, and overlapping joint discs so a
  turned joint shows no gap. The person is a collage, as the method does it: a face in a hat (Vision of the Wild f010,
  flopped to face the dog) with its chin cut free and pinned at the ear over a scrap of black paper (the hinged jaw), a
  smock (Straight Talk f000), jeans cut at the knee (Straight Talk f013), a forearm that held a card (Door to Heaven f000).
- **Camera frame interval**: `xsheet.json` has 255 exposures for 768 frames. Every exposure is held at least 2 frames
  (on twos), re-phased so each choice starts its own exposure, and held longer whenever nothing on the table is touched
  (longest hold 24 frames; 255 exposures after the repair round). A card that is touched is never put back exactly (a small placement error per exposure on
  the head, tail and person); untouched cards do not move. Film grain changes every frame; the lamp flickers a little
  per exposure.
- **Shot breakdown**: one locked-off rostrum shot (the camera does not move), the facade face-on, the porch running along
  it; space is `x_screen = 290 + 210 · x_metres` (threshold at the door centre, the door opening at ±0.46 m = 193..387 px).
- **Rights advisory**: `assets.json`; nothing of unknown rights is used; Quick, Draw! was not used; no web image, no
  model-generated image.
- **Reconstruction drawings**: `proof/cutting_sheet.jpg`.

## How the dog moves (the operation)

- **Walking**: the leg cards swivel on their pins. Each foot's offset from its pin is linear in the distance travelled
  while it is down, so the foot stays where it was put on the boards; from a standing start each foot stays where it stood
  until its first swing. Measured from the placed transforms: 270 consecutive-exposure pairs of a planted foot, mean slip
  0.43 px, max 1.89 px (the leg prints are slightly slanted, the stance law assumes vertical).
- **Backing away (1.00 s)**: the same legs run backwards, head card dropped, tail tucked, body tilted down at the rear.
- **Turning (3.42 s and 8.08 s)**: never a squash or a mirror sweep. **Substitution**: the profile is lifted off and a
  second print of the same dog, facing the camera with its head turning, is laid in its place for two exposures; then the
  **flopped** puppet (a mirrored print, because the back of a photograph is white paper). Every frame of both turns is a
  whole photographed dog.
- **The refusal**: back to the door from 3.42 s (measured facing −1 on 100% of refusal frames), then it lies down facing
  away, chin on paws.
- **Turning to look (7.458 s)**: the head card alone is substituted, on the choice frame, by a face card (the same dog's
  face from f030) looking left at the door, while the body still lies facing away. It then rises and turns by the
  turning card.
- **Coming, head low (14.708 s)**: the head card hangs from a lower pin on the chest and is swivelled down; measured, the
  topmost pixel of the head card is below the withers on all 219 approach frames (min 25 px), the muzzle 46+ px below.
- **Lying down across the sill (23.792 s)**: the archive dog was never filmed lying. Solved in the language: the legs
  swivel flat (forelegs forward along the boards, hind legs folded), the body card comes down onto the boards, rear not
  raised. It ends with 50% of its pixels inside the door frame on screen (head and shoulders in the lit doorway, hindquarters
  and tail outside), on the porch plane: it never goes through.
- **Looking in (26.4 s)**: from 26.42 s the head card is substituted by a print of the back of the same dog's head
  (f001, round 4 repair), level, ears hanging; from 28.08 it lowers toward the sill and stays turned in. Then the
  profile head comes back and lowers, chin on paws, and stays.

The person: calls with the hinged jaw opening on the voice (the jaw is driven by a mouth chart read from the voice takes,
as an animator reads a track), reaches with the forearm card, steps aside, sits on the crate (thighs swivel at the hip
pin), offers the hand low and forward, and nods. The door is a **substitution set** of five prints of a real door
swinging (Straight Talk f024 → f022 → f020 → a strip of its edge): found motion re-timed and reversed. Light coming out is
the table lamp: a warm lamp in the room whose light falls on the boards and the dog through as much of the doorway as the
door uncovers.

## Protected conditions and invariants

Numerical (measured, `proof/measured.json`, `invariants.json`):

- PAELEO check: `check_handmade.py` ALL OK (I1 choices, I2 never inside / ends on the sill, I3 direction, I4 refusal).
- **Onsets measured from the render**, not from labels (trace.json is measured from the rendered dog layer each exposure):
  retreat 1.000 s (head drop / backward travel), stops and head comes up 2.333, back turned 3.417, head turned to the
  door 7.458, approach 14.708 (head drops and travel begins), lies down 23.792 (body top drops), looks in 26.417. All within
  0.02 s of the score.
- **Rigidity** (structural, asserted in `table.place`): every placement is a rotation + translation (+ mirror), |det| = 1;
  no piece is ever scaled, bent or warped after it is cut.
- **Camera interval**: every exposure ≥ 2 frames (100%).
- **Sound** (measured on the synthesized stems): calls −18.3 dBFS active RMS, soft voice −31.4 (13.2 dB quieter than the
  calls), room tone −46.9 (soft voice 15.5 dB above it). `proof/sound_rms_dbfs.png` is the track's level over time.

Judged only by viewing (not by me): whether the choices read as choices, whether the collage is legible and has the
expressive roughness the method promises, whether the dog reads as a dog in every frame.

## The sound

Synthesized. Voice: flite speech synthesis (ffmpeg's libflite, voice "slt"), two calls and three soft lines; each line's
first sound starts 0.15–0.2 s before the dog's choice so the cause is heard before the effect. Room tone of dusk air and
crickets; claws on boards placed at each foot-fall of the gait (from the exposure sheet); a whimper as it turns its back;
breaths as it lies down and settles; a dry tap of card when a card is swapped (the turns, the head cards); a door creak;
a reed-organ chord when the lamp's light comes out, and another when it lies across the sill. **I cannot listen**: levels
and timing are measured; whether it sounds good is unknown. The synthetic voice is plainly synthetic.

## Repair round 1 (independent critic: does not pass, near)

The critic's five conditions and what changed (protected: the turns, the refusal, the lie-down, the look-in and settle,
the sound levels):

1. **Head low, wholly below the withers (14.71–23.79 s).** My earlier measure used the head card's centroid; the critic saw
   its dark ear and skull at the line. Now measured as the critic asked: the topmost pixel of the head card (ears included)
   against the top of the torso card in the column of the shoulder pin. The head card now hangs from lower on the chest
   (a lower pin, 30 px) and is swivelled down 60°. Measured on all 219 frames: topmost head pixel 11–15 px below the withers
   (min 11), muzzle 60–67 px below.
2. **One unambiguous face at the door by 7.70 s.** The turned-over profile head is replaced, from 7.458 s, by a new
   substitution card: the face of the same dog cut from f030 (eyes and muzzle readable), laid looking left on the
   lying, facing-away body; it stays through the rise until the turning card at 8.08 s. Measured: 100% of frames
   7.458–8.1 s have the head toward the door.
3. **No hyena blotches, no jagged hackles.** All dog prints regraded (the print is blurred 1.7 px and its contrast
   compressed to 0.62 before the dye ramp, no sharpening, a calmer palette) and the scissor line follows a smoothed
   matte (blurred 2.2 px and re-cut at half), so the back edge and tail are smooth. Looked at in full-resolution crops.
4. **A deliberate absurd beat.** 11.9–13.7 s, while the dog only watches: a giant hand (a second, much larger print of
   the same forearm) slides in from the edge of the table, takes the moon between finger and thumb, lowers it into the
   cloud (the dusk deepened by hand), and withdraws, with paper swishes and a descending slide whistle. The dog and the
   person are not touched; I1–I4 unchanged.
5. **The back of the head reads as a dog's.** Rebuilt: a broad skull on a thick neck, upright broad-based ears with the
   top third folded outward (darker fur), and the tip of the muzzle (a scrap of f046, flopped) showing past the skull on
   the side it is turned to, inside the door frame.

## Repair round 2 (critic: pass, narrow; one more round)

1. **Look-in, 26.42–28.08 s.** The back-of-head card is rebuilt again: long hanging ears down the sides of the skull, like
   the profile head's (the pricked ears read as a cat's), and the muzzle tip (a scrap of f046) showing over the top of
   the skull, pointing up and to the right into the lit room, away from the person and the hand. During the hold the card
   tilts toward the room at 26.95 s (+6°), back past level at 27.45 s (−3°), and settles at 27.8 s (+1°), so the look is
   not frozen.
2. **Approach reads as a dog.** The torso card's back line is cut lower (no withers hump) and its matte smoothed harder
   (4 px), the whole body is levelled 4° (the print's back sloped down to the rear: the hyena), and the head hangs from a
   pin 46 px down the chest, swivelled only 32° so the muzzle reaches forward and down rather than drooping (the tapir).
   Measured on all 219 approach frames: topmost head pixel 25–29 px below the withers, muzzle 46–54 px below.
3. **Face card at 7.46–8.04 s.** Re-cut from f030 at a smaller print scale (0.74 instead of 1.31): 39 px tall against the
   torso card's 68 px (ratio 0.57, required ≤ 0.6). It still replaces the head on the choice frame 7.458.
4. (optional) **Tail tucked** between the hind legs through the refusal (pinned at 70° standing, 58° lying).
5. (optional) **Footfalls audible**: claw-and-board clicks raised 9 dB; 13 transients more than 12 dB over the bed during
   the approach, plus a paw and a body thud at 23.84–23.9 s (20 dB over the bed). Measured, not listened to.

## Repair round 3 (critic: does not pass) - by a second builder

The first builder left before finishing this round; a second agent continued from its files (its unfinished edits: legs
cut in two at wrist and hock, eyes and noses inked on the head and the turning card). All times are scene time.

1. **The look back, 7.458-8.08.** The look-back card is a second print of the f006 profile head at 1.12 (was 0.82, which
   became a featureless lump under the regrade): 63 px tall, 0.80 of the torso card, kept sharp, eye, nostril, mouth line
   and ear retouched in ink. It is laid on the choice frame with the muzzle level at the door. Measured from the placed
   transform: muzzle direction 169 deg (180 = level toward the door); 100% of frames 7.458-8.08 face the door. Looked at
   in film frames at 7.46, 7.58, 7.71, 7.92, 8.04 (2x crops): a face with an eye and a nose turned back over the shoulder.
   This deliberately undoes round 2's size condition (<= 0.6), which round 3 found had cost the features.
2. **The front card (3.42-3.54, 8.08-8.29).** The f030 turning card kept sharp and retouched (two eyes, nose, ear edges,
   a lighter muzzle wash). Looked at in film frames 3.42, 3.50, 8.12, 8.25: a dog's face and forelegs.
3. **Looking in, 26.42-28.04.** A new constructed card (`dog_away`), a lost profile: the inked f006 head card laid first,
   and over it the back of the skull, the nape and both ears cut from the f006 saddle fur, hiding the eye and the cheek,
   so only the muzzle tip and nose show past the skull. It is laid *under* the torso card, so the nape rises out of the
   shoulders, and **flopped**: unflopped, its muzzle pointed straight at the offered hand (tried, rejected). Flopped, the
   muzzle points 24 deg above level to the right, into the lit doorway, away from the person and the hand, and well
   below the lamp (about 95 deg). A warm rim of room light is painted on the side facing the room. Measured: muzzle tip
   inside the door opening (x 193-387) and above the sill on all 40 frames. During the hold the card reaches a little
   further in at 26.95, eases back at 27.45 and settles at 27.8 (4-7 deg). Judged by eye only on film frames; whether a
   cold viewer reads it as "into the room" is the critic's call.
4. **The watch, 8.33-14.71.** The ear is now its own card, cut from the same print and pinned at its root. Acting beats
   (from the exposure sheet, all visible in film-frame crops): 8.75 the ear pricks; 9.67 head and tail settle; 10.75-11.9
   the listening cock of the head at the second "It's all right", ear dropped forward; 12.5-14.0 the head lowers and the
   ear relaxes (deciding); 14.0-14.5 a forepaw lifts, hanging from the bent wrist, and is set down. The approach onset
   is unchanged (14.708).
5. **Folded legs lying (4.5-7.4, 24.2-32).** Each leg is two cards. Lying, forelegs 60 deg on the shoulder pin with the
   lower card level (90) and **flopped** (mirrored across its own axis) so the paw lies toes-down instead of toes-up (the
   flipper); hind legs 95/90, flopped: the haunch forward with the hind paw under it. Toes inked on all four paws.
   Rejected: rotating the lower card past 90 (it lifts like a boot); folding the hind leg backward (it goes through the
   boards). Measured: forepaws within 6 px of the board line; the hind paw lies 14-18 px above it, under the haunch.

Also this round: `build/assets.py` regenerates `assets.json` from the pieces as cut (it had fallen behind the cutter);
the unused f046 back-of-head card is no longer cut and its source copy is removed; the end card now says twelve pieces
and declares the ink retouching as drawn here. Kept unchanged: timing of all choices, the head-low approach (topmost
head pixel 25-29 px below the withers on all 219 frames), the lie-down across the sill, the moon beat, the sound
(soft voice 13.2 dB under the calls, 15.5 dB over room tone, voices before the reactions).

Still weak (my view): the ear card's cut edge makes a slightly boxy outline on the lying head; the hind fold reads as a
haunch more than as a leg; the look-in's muzzle tip is small at full frame (about 15 px); the rise from lying (7.5-8.0)
passes through half-folded leg cards for a few exposures.

## Repair round 4 (critic: does not pass, on the look-in only)

The round-3 lost-profile card read at full frame as a head lifted upward, its muzzle tip above the ear line. Replaced:
the look-in head is now a **real print of the back of this dog's head**: Amateur Film New Jersey C f001 (public domain),
where the dog is seen from behind looking at the person. It is scissor-cut at the dog's own print scale (1.31) to a
crown, two ears hanging evenly and a short nape (the print's ears are lost in its shadow, so the ear flaps and the
parting of the fur are retouched in ink), and laid under the torso card at the nape, level. No muzzle shows. Measured
from the placed transforms: crown at x 333.5 (inside the opening 193-387) on the choice frame 26.417; the card's angle
during the hold is 0, +5, -4, 0 deg (26.95, 27.45, 27.8); from 28.08 it slides 12 px down (crown y 495 -> 507), chin
toward the sill, and stays: it never swaps back to the profile facing the hand. 134 frames (26.42-31.96). The
constructed lost-profile card is no longer cut. Looked at in film frames, full frame and 2x (26.42, 27.0, 27.6, 28.0,
28.5, 31.95). Nothing else changed. The profile head may still read as an otter (optional condition, not done).

## What failed and what I changed (see ledger.jsonl)

- First head cut took the whole mane: lying, the dog read as a seal. Cut the head smaller at the neck line.
- A chest point of the torso card poked below the boards when lying (a triangle under the chest). Re-cut the chest.
- The face print looks down and to the left, not right: the head and jaw are flopped on the table, the jaw re-cut with
  the hinge at the ear.
- Feet slid up to 10 px at the start of the approach (blending into the gait). Replaced by a planted start: each foot
  stays where it stood until its first swing. Max slip now 1.9 px.
- The trace's facing first measured head position against the torso, so the look back at 7.458 read as 8.08. Facing is
  now measured as the direction the head card points from the neck pin.
- The voice takes had 0.25 s of leading silence: the dog backed away before the call was heard. Trimmed and led.
- Lying bodies were tilted rear-up 3°; levelled. The low hand was too low to read as offered; extended.
- Disk full mid-encode (shared machine); the film is a two-pass 2.1 Mb/s encode (10.7 MB).

## What is still weak (my view, not a verdict)

- The dog is small in a wide, locked-off shot (about 210 px long). The regrade removed the spots but also most of the
  photographic detail: the dog now reads as a smooth tinted card more than as a photograph.
- The lowered head hangs from the chest: below the withers as required, but the neck can read long.
- The giant hand is in the sky, away from the dog; whether it reads as absurd or as a distraction is for the viewer.
- The half-way exposures of lying down and getting up (legs part-swivelled) look like a sprawl for two or three
  exposures.
- The person is crude: a box-edged hat card, wide smock, the face mostly in shadow; the closed jaw shows its cut as a line.
- Rigid single-piece legs cannot bend, so the crouch is mild.
- The voice is synthetic; music and effects are synthetic.

## What I did not do

No physical paper, no filmed tabletop test, no access to the broadcast or any Gilliam material, no independent critique
(that is for the critic), no camera moves, no on-screen typed prompts (the voice carries the prompts), no Quick, Draw!,
no publishing of an operator page outside this folder (a local `index.html` is here).

## Rebuild

`build/make.sh` (cut pieces → voice → exposure sheet → render + trace → cards → sound → measure → stills → encode →
check). Intermediates in `build/cache/` (regenerable; scene frames are deleted after encoding to keep the shared disk lean).