# PAE-02 · SCION 03 · Max Ernst — Surreal Collage Grammar · method ## Thesis One plate from a collage-novel: a porch at dusk, engraved. Its impossible things are fixed in place and its surfaces are continuous: one burin cuts every part, so you cannot see where the lion, the hand or the lamp were pasted on unless the film wants you to. The defining passage is the lion. While the person gives orders its head is a lion, cut from a film and pasted on with its white scissor margin showing. When the person steps aside, the lion leaves the head and settles over the door as its keystone. There the margin fades: the authority has become architecture. Then the door opens on its own. At the end the dog lies across the sill below the lion, the way the lion lay in its film. The motif comes back three times, each time doing a different job: commanding head, keystone, the dog lying couchant at the threshold. The film would be bland if the lion were only decoration, if the collage were a random pile, or if the dog were a flat cut-out that flips. ## The swarm (merged roles, one builder) The SCION prompt names six roles. Here one agent did the building, and the work was split into artifacts each role owns: | role | artifact | |---|---| | archive iconographer | `build/fragments.py`: it picks the found frames, cuts them by their SAM masks, engraves them and copies the sources into `src/`. `build/assets.py` writes the rights ledger (`assets.json`) | | hybrid form composer | `build/dog.py`, `build/person.py`: the maquettes. The person's lion head and pasted hand are composed in `build/render.py` | | seam/mat specialist | `build/burin.py`, `build/dogplate.py`: one line language. It also sets the seam rule: margin shown means pasted (foreign), margin hidden means absorbed | | recurrence dramaturg | `build/timeline.py`: what the lion and the hand do, and when | | improbable-world builder | `build/plate.py`: the porch, the door, the light coming out, the dusk | | independent critic | **not me**. I did not approve this film. The critic sees the film, the brief and the bar | ## Native formal language, as built - **Engraving source fragments.** There are five, cut from three public-domain films: two 3/4 lion heads and a frontal lion head (*Andy's Animal Alphabet*), a hand (*The Door to Heaven*), and a miner's safety lamp (*Anthracite: Pre-Shift Examinations of Underground Mines*). Each is re-engraved in this plate's burin. The lion gets a radial line system, so the mane reads as engraved fur. - **Hybrid nouns.** The person-lion is the magistrate. The hand-that-swells is the reach. The lion-keystone is authority turned into architecture. - **Transformation rules.** - Command → the lion head is on and its margin shows. - Stepping aside → the lion flies to the keystone and is swapped from the 3/4 head to the frontal head in mid-flight. - Landing (a stone knock) → the door opens on its own and light comes out. - Reach → the pasted hand swells to 3.4× along the arm. - Offer → the same hand at normal size, held low, with no margin. - **Compositional seams.** These follow a rule rather than being left to chance. A fragment acting as an intrusion keeps its white cut margin. A fragment that has been absorbed loses it. Everything else (dog, person, set) is cut by the same burin, so those seams are hidden. - **Recurring icons and symbolic motifs.** The lion (authority → threshold → the couchant dog), the hand (grab → offer), and the light (a sliver through the door left ajar → a hall lit by the lamp). - **Rights ledger.** `assets.json` lists every fragment with its file, source title, archive URL, clip id, rights and operation. ## How the dog is built The dog uses no picture of a dog. It is a solid maquette, like an anatomical-plate dog, made of 27 typed primitives (round cones and ellipsoids): - chest, loin and haunch - neck, skull, muzzle and nose - two eyes and two ears - four legs of three segments each - a tail of four segments It is posed by typed parameters (yaw, lie, crouch, head, rest, ear, tail, wag, gaze yaw and gaze pitch) and built at unit size, then scaled ×1.2. It is ray-marched in the plate's orthographic camera (`build/sdf.py`). The burin then cuts lines that run *along the form*: the angle around each primitive's axis sets a line count that wraps without a seam. Rings cross-hatch the deep shadows. The coat has a darker saddle, the dog is lit from the sky above and by the warm light from the door, and it is outlined at its silhouette and at depth breaks. Because the lines are attached to the body, they do not boil. ## How the dog moves - **Position.** `x` comes straight from the score: screen x = 400 + 168·x, with a fixed camera. - **Turns.** - The yaw always passes through 90°, so the dog shows its face to the viewer mid-turn. A solid turning in 3D never goes edge-on. - The refusal turn runs from 3.42 to 3.92 s, so the back is shown by about 3.9 s. - At the turn back (7.46 s) the head comes round first, over the shoulder on the viewer's side, and points at the door by about 7.7 s. The body follows by 7.96 s while it rises. - **Feet** (`build/gait.py`). A planted paw is a constraint and stays put in world coordinates. The paws are lifted in the lateral sequence of a walk (LH, LF, RH, RF). The gait clock advances with distance travelled (one stride per 0.54 m, the same rate as the score's walk phase) and with rotation. Each paw lands where the body will want it half a stance later, since the solver looks ahead along the known timeline. In a turn on the spot the paws step quickly, one per side at a time. - **Head low.** During the approach the head is lowered 0.13 below the score's head value, and the measurement says it stays below the line of the back. - **Lying across the sill.** For the last 0.6 m of the approach the dog angles 27° into the doorway (z from 0.62 to 0.10). It lies with its forequarters in the door opening and its tail out on the boards. This was solved in the maquette's own language, since there is no lying dog in any of the kit footage. - **Looking in.** At 26.4 s the gaze target becomes the hall inside (score `look`), and the head turns about 60° into the lit opening. Then the chin goes down onto the paws and the dog stays. ## The invariants (measured: see `invariants.json` and `trace.json`) `trace.json` is measured from the render. In each frame: - `x_screen` and `y_screen` are the centroid of the dog's rendered torso pixels. - `facing` compares the rendered head and tail on screen. - `posture` comes from the rendered torso height. - `head_below_back`, `frac_in_frame` and `z_min` are also measured. `state` is the score decision that drove the frame. It is copied, not measured, and that is said here plainly. - **I1 choices**: all six found at dt = 0.000 s. This is structural: `state` is the score's decision for each frame. - **I2 threshold**: min x_score 0.59 and end 0.59. Visually, the dog's deepest rendered point is z = -0.222 m, inside the door frame's reveal (which runs from 0 to -0.24 m) and short of the door, which hangs at the inner face. - **I3 direction**: r = 1.0 on x_screen, measured from the rendered torso pixels with a fixed camera. - **I4 refusal**: share turned = 0.938. Measured facing first flips at 3.667 s; the turn runs 3.42 to 3.92 s. - **Head low on the approach**: the rendered head centre is at least 8 cm below the top of the back in 99.1% of approach frames. - **Lying across the sill**: posture is 'lying' (rendered torso height under 0.29 m) in every frame from 24.5 s (179/180 frames). At the end, 49% of the dog's pixels lie inside the door opening's x-span (head and forequarters) and the rest outside (hindquarters, tail). - **Turn back**: the head's relative yaw reaches 111° at 7.67 s, so the head comes round over the shoulder before the body turns. The rendered head is on the door side of the tail (measured facing +1) from 7.71 s. - **Soft voice vs calls** (measured on the synthesized voice track): calls -18.9 / -16.1 dB RMS, soft voice -29.9 dB, so the soft voice is 11.0 dB under the quieter call. Room tone is -45.1 dB, so the soft voice is 15.1 dB above it. The final mix needed no gain change (peak 0.627). - **Looked at** (by me, on extracted frames): the turns, the walk, the lie-down, the look-in, the lion's transfer, the retreat, the four stills and a 1 fps sheet of the whole film. Proof sheets are in `proof/`. - **Unknown**: how it plays at normal speed to a fresh eye, and how it sounds. ## Sound Every sound is synthesized (`build/sound.py`); nothing is recorded. - The calls are formant voices ("Come here." / "Here!") with a lion's growl pasted under them. - The soft voice ("It's all right.", "Come in.") is breathy and lower. - Paper rustles as the hand swells and as the lion flies. A stone knocks when the lion lands. - The door gives a hinge creak, then a warm low chord: the light, heard. - The dog makes a claw tick and a pad sound at every touchdown the step solver reports. It sighs when it lies down, and its tail knocks the boards while it looks in. Measured levels are in `build/cache/audio_levels.json`, as RMS over the voiced samples (summarized below). **I cannot listen.** Whether the voices are intelligible as words, or merely voice-like, is unknown to me. I checked only spectrograms, which show harmonic, formant-shaped utterances where the voices should be. ## What failed, and what changed (details in `ledger.jsonl`) Each line of `ledger.jsonl` is one attempt with its evidence. In short: - **The maquette** first read as a weasel, then a deer. I changed its proportions, legs, head and ears, and scaled it ×1.2. - **Lying parallel to the facade** made looking in read as looking up. Fix: the dog angles into the doorway at the sill. - **The 3/4 lion** read as a sculpted human face. Fix: a radial, mane-like line system. - **The lantern** on the hall's back wall showed above the door, because of the orthographic camera. Fix: it now hangs in the hall. - **The gait** lurched in trot pairs. Fix: a lateral-sequence walk on a distance clock with look-ahead landing. - **The refusal turn** left the legs stretched. Fix: a fast turn mode for the steps. - **The turn back**: the chin-on-paws pose held the head, so it came round late (about 8.1 s). Fix: release that pose and let the head lead. - **The growl** under the calls was broadband because of a phase bug. Fixed. - **The offered hand** sat on a stretched stick of an arm. Fix: the reach is clamped, and the person leans forward with the hand low. - **The nose** went 0.25 m into the wall, past where the door hung. Fix: the door now hangs at the inner face, and the dog's depth was reduced. - **Looking in** rose too steeply, toward the keystone. Fix: the gaze target is now low and deep in the hall, and the head is lowered. - **Rendering** was too slow. Fix: a 2D cull of the rays. The output was verified identical. What is still weak: - The look-in still points up-left on screen. In this camera, depth reads as up. A viewer may read it as looking up at the lamp or the lion rather than in. - The turn at 3.6-3.7 s passes through a small frontal view that is barely readable as a dog at plate scale. - The 3/4 lion head is ambiguous between lion and carved face. - The legs are stiff, like stilts. - The person is a dark, simple mannequin. - The six seconds of waiting (8.3-14.7 s) are nearly static apart from ear flicks. - The voices are synthetic. Whether they are intelligible is unknown. ## What I did not do - No independent critique has happened yet. The verdict belongs to the critic. - Not used: the kit's archive dog, the archive door, and Quick, Draw!. I looked at all three; the dog is a maquette instead. - No motif graph page or operator page; the deliverables of this brief were followed instead. - No forked branches were compared. - Sound was not listened to. ## Round 2 (after critic round 1: PASS, narrow) The critic's largest gap was that the collage grammar stopped at the dog: a self-built 3D model with a hatching shader. I made one repair round and kept what the critic protected: the lion becoming the keystone, the hand changing job, the 3D turns, the head-led turn back, the look-in, and the sound mix. 1. **The dog now carries the grammar: a lion's mane round its neck.** - **Source.** The mane comes from the same found frame as the magistrate's head (*Andy's Animal Alphabet*, lion f022). Its face is cut away with a disc, leaving a crescent of radially engraved mane. It is mirrored when the dog faces away (`mane`, `mane_r` in `assets.json`). - **Placement.** The hole sits on the dog's skull, and the dog's head is re-composited in front of it. The result is a dog wearing a lion's ruff. - **Its job changes.** - At the first call (1.0 s) it is pasted on with its white cut margin: hackles up. The command's lion has got into the dog. It stays through the refusal. - It shrinks as trust grows (7.6–23 s). - As the dog lies across the sill (23.9–24.9 s) the margin fades to nothing and the ruff sinks into the coat (82% opacity, paper tint). It is absorbed as the ruff of a couchant guardian lying under the keystone lion. - **What it means.** The dog's change from refusal to acceptance is also a change in what it is made of: from a pasted intrusion to its own coat. The seam rule is the same one the lion and the hand obey. The per-frame scale, margin and opacity are in `trace.json` (`mane`). 2. **Head low.** During the approach the head is lowered 0.27 below the score's head value (was 0.13). At 16, 19 and 22 s the rendered head centre is 0.48–0.49 m and the top of the back is 0.78–0.79 m. The skull, whose vertical radius is about 0.1 m at scale, sits about 0.2 m below the back line. The head is below the back in 99.5% of approach frames. 3. **A canine head.** - The muzzle is longer, with a stop: it now reaches 0.205 at unit scale, up from 0.152, and is narrower. - The skull is flatter. - The ears are larger (0.068 × 0.04) and never fully flat; pinned back, they still stand out from the skull. - I looked at 1.9, 6.0 and 29.5 s (`proof/r2_critic_times.png`). The head reads as a dog's (hound or wolf-like), though at plate scale it is still small. 4. **Captions.** - A caption now shows when its circumstance is set, for 2.4 s, and again while its words are spoken (0.8 s before to 0.5 s after each voice span). It fades over 8 frames. - The 12.1–25.8 s span and every other silence are now without a caption. 5. **The door is struck open by the lion.** - The lion lands in the keystone at 6.05 s with a visible drop and rebound and engraved impact strokes under it (6.05–6.32 s), together with the stone knock in the sound. - The door starts to swing on that knock (6.07 s), fast at first, and the light rises with it from 6.1 s. - The score's circumstance "door open" is 6.2 s. The swing starts 0.13 s earlier so that the knock is its cause. **Trace.** `trace.json` was rewritten from the new render. It now carries `measured_onsets`: when the rendered dog visibly does each choice, measured from its own render (pixels, rendered heights), not from the `state` field. | choice (render measure) | score t | measured onset | dt | |---|---|---|---| | retreat (backs away: x_screen rises 2 px) | 1.0 | 1.083 | +0.083 | | watch (stops backing: |v| < 25 px/s) | 2.333 | 2.583 | +0.250 | | refuse (rendered head passes to the far side of the tail) | 3.417 | 3.667 | +0.250 | | watch (head starts coming round: head_x moves 3 px toward the door) | 7.458 | 7.5 | +0.042 | | approach (comes: x_screen falls 2 px) | 14.708 | 14.792 | +0.084 | | lie_down (rendered torso drops 3 cm) | 23.792 | 23.833 | +0.041 | How to read the two +0.25 s values: - The refusal turn starts at 3.42 s; 3.67 s is when the rendered head crosses to the far side of the tail. - The retreat's "stop backing" lags because the score's own x is still decelerating after 2.33 s. `check_handmade.py` reports ALL OK (I1–I4). The film is 38.0 s and 4.24 MB. **Round 2 also changed.** The colophon now names the dog's mane and the lion's three jobs. Its rights line names Literata (OFL) and no longer says "everything else made here" about the type. **Still weak after round 2:** - The ruff at 24.9 s and later is subtle at plate scale. Whether the viewer sees the absorption, and not just a disappearance, is for the critic to judge. - The person still lurches to the chair (5.4–6.15 s). - The legs are still stiff. - The look-in still points up-left on screen. - I did not listen to the sound. It is unchanged apart from the new door timing; the knock was already at 6.04 s.