SCION

SCION

Fifty-four animation swarms. Each is a standalone prompt that turns one artist, researcher, studio, algorithm or archive into a working studio with a proof to make and a critic to answer to.

The SCION operator is a single page: search the 54, read one, copy it whole or save it as Markdown. Each prompt names a lineage, states a creative theorem, gives the swarm its native formal language and the smallest crew that could build it, and sets a hard proof scene to produce before anything else. Then all 54 inherit the same fifteen-point doctrine. PAELEO films are made with them, one swarm per box.

The reel's critic rounds

Reel round 1

does not passa fresh critic · a mid-frame of every lens, dog crops, cut detection, sound by measurement

Largest gap. The formal language sat next to the dog instead of on it: about a dozen lenses read as the same brown dog with a new caption at phone size.

Repair condition. At 360 px wide, at least 25% of the dog's pixels differ visibly from the plain dog in every lens; no two lenses within three places share an overlay type.

Weak: 06, 10, 13, 18, 26, 32, 33, 40, 46, 49; near-duplicates 07/38, 25/48, 12/36/37.
Collisions: 14, 27, 47, 49 over the feet or the porch.
No rig pops; the 1.6 s grid exact; strong: 02, 07, 08, 17, 24, 29, 31, 36, 51, 53, 54.

What was done. Fifteen lenses redrawn so their language is on the dog (harmonic points, a warped lattice, follow-through ghosts, gilded spirals, sprite tiles, angle arcs, solver-only IK, rubber warp, red-pencil cards, a tilted tabletop, a museum seamless, a face mesh, waveforms through the body); overlays moved off the action; the critic's own pixel test, run by script, passes for all 54 (lowest 28%).

The critic's report, verbatim

REEL CRITIC, round 1: paeleo-01.mp4

What I inspected: ffprobe; contact sheets every 0.5 s for the whole film; one full-res mid-frame for each of the 54 lenses, tiled with numbers; enlarged dog crops for lenses 6, 9, 10, 13, 18, 26, 32, 33, 40 and 46; four frames around 28.1 s; cut detection by frame difference over the whole frame and over the dog region; audio RMS every 0.25 s; a full spectrogram. I did not check each lens frame by frame, and I did not judge caption type size on a real phone.

Timing: title 0–4.0 s, lenses every 1.6 s from 4.0 s (lens n starts at 4.0+1.6(n−1)), end card 90.4–95.4 s. The structure is exact.

1. Verdict: does not pass. The rig and the structure hold up. But about a dozen lenses make their look visible only as tiny dots, faint lines or a corner panel. At phone size they read as the same brown dog with a new caption, so the bar of "visibly distinct and legible in 1.6 s" fails for roughly 20% of the reel. Several neighbouring or repeated lenses also look alike.

2. Weakest lenses:
- SCION 06 Whitney, 12.0 s: "harmonic motion" is a few pink and purple dots and faint arcs at the paws.
- SCION 13 Willett, 23.2 s: two red dots on the nose and tail. Nothing reads as secondary motion.
- SCION 10 Schaefer et al., 18.4 s: faint teal vertical lines, which read as rain or glitch. SCION 35 (58.4 s) is also "moving least squares", so the method appears twice.
- SCION 18 Cartoon Saloon, 31.2 s: only a faint amber frame edge. The dog is untouched.
- SCION 26 PixiJS, 44.0 s: a stats box in the corner. The dog is unchanged.
- SCION 32 FK (53.6 s) and SCION 33 IK (55.2 s): thin orange bones, then red feet with white contact ticks. Cut detection barely registers 33 to 34 (56.8 s), so the change is very small.
- SCION 40 Physical tabletop, 66.4 s: just a darker grade and a soft shadow.
- SCION 49 Getty, 80.8 s: a label card only. The dog is unchanged.
- SCION 46 Face landmarks, 76.0 s: the dots sit mainly on the person's face. On the dog it is a small cyan cluster at the snout.
- Repeated looks: 07 and 38 (cream line drawing); 25 Rive and 48 State machines (the same watch/approach/retreat/refuse panel); 12, 36 and 37 (weight colouring).

3. Broken picture or collisions: I found no rig pops. The only jump the dog-region check flagged (SCION 16 at 28.17 s) is a normal leg change in the walk cycle. Collisions:
- SCION 27 Remotion (45.6 s): the timeline bar covers the person's legs and the dog's feet.
- SCION 14 Willett (24.8 s): the pose filmstrip runs across the person's feet and the porch line, and is tiny.
- SCION 47 (77.6 s): the waveform box sits on the porch next to the dog's paws.
- SCION 49 (80.8 s): the card crosses the porch edge under the dog.
- SCION 24 Live2D (40.8 s): a sudden close-up. Effective, but it is the only scale change in the reel and breaks continuity with 23 and 25.

4. Largest gap: the formal language sits next to the dog instead of on it. Repair condition: in each lens's mid-frame scaled to 360 px wide, at least 25% of the dog's pixels differ visibly from the plain dog. Measure as the mean absolute difference in the dog box against a reference plain-dog frame of the same beat, and set a threshold. In addition, no two lenses within 3 positions of each other may share the same overlay type. Today 06, 13, 18, 26, 32, 40 and 49 would fail the first test.

5. Strengths to keep:
- The 1.6 s grid is exact.
- A broadband click on every cut, a quiet drone bed around −35 dB, and voice-like call bursts near −13 dB on the "call" beats. The sound marks the rhythm cleanly.
- Strong lenses: 02 Reiniger, 07 Foldès, 08 Sutherland, 17 South Park, 24 Live2D, 29 GrabCut, 31 Alpha matting, 36 BBW, 51 Mouris trails, 53 Atlanta film strip, 54 Provenance graph.
- Caption bar (number, family, name, look) is consistent and doesn't collide.
- The dog rig is recognisable and continuous in every lens.
- Title and end cards are clean.

Reel round 2

does not passa second, fresh critic · mid-frames of all 54, dog crops at 360 px, five frames across six lenses, cut onsets

Largest gap. Overlays repeated as the way to show a method, where the method should be on the dog: five lenses (05, 24, 40, 44, 53) still left the dog plain, and neighbours shared corner tags and dark panels.

Repair condition. Every lens changes at least 25% of the dog at 360 px; no two lenses within three places share an overlay type; lens 14's dog keeps at least half its area in every frame.

Breaks: 14's dog flipped to a sliver; 38 pasted a card over the dog; 49 removed the person.
25 and 48 both state graphs, and 14 used the same state words.
Captions: the look too small; two "NORA WILLETT" titles; the cut into 17 and the end card had no sound onset.

What was done. The five plain dogs get their method on the body (glass planes, a deforming mesh, card thickness and a lamp, depth-tinted planes, silver-gelatin with scratches); corner tags and panels removed or moved onto the dog; 48 shows trust as a level inside the dog and its last reason, 25 a graph tinted by state; 14 at a moment with no turn and numbered drawings; 38 keeps the dog under the red pencil; 49 keeps the person, greyed and out of focus; captions carry the full title and a larger look; a click on every cut and the end card.

The critic's report, verbatim

REEL CRITIC, round 2: paeleo-01.mp4

What I inspected: the mid-frame of all 54 lenses (at 640 px, plus dog crops scaled to 360 px wide); 5 frames spread across each of lenses 14, 24, 27, 38, 47 and 49; the title at 1.5 s and the end card at 92 s; three plain-dog reference frames from paeleo-00.mp4; and the audio, by measuring loudness (RMS) and looking for sound onsets at every cut time. I did not watch it in motion beyond those samples, and I did not build a pixel mask of the dog, so the 25% test was judged by eye against the paeleo-00 dog.

1. Verdict: does not pass. Most lenses now change the dog clearly, but both parts of the previous critic's repair condition still fail. Four lenses leave the dog essentially plain, and several neighbours within 3 positions share the same kind of overlay.

2. Lenses that still fail:
- 05 Norstein (10.4–12.0 s): the dog is a slightly washed-out plain tan; the multiplane look is only foreground blobs.
- 24 Live2D (40.8–42.4 s): a close-up where the body stays plain tan cut-paper; only the head has a cone and a pink glow. Well under 25% of the dog changes.
- 40 Physical tabletop (66.4–68.0 s): the whole scene is skewed, but the dog itself is plain.
- 44 Multiplane parallax (72.8–74.4 s): plain dog behind foreground bushes. It also repeats 05's idea.
- 53 Atlanta Moving Image (87.2–88.8 s): the sepia grade covers the whole frame; nothing is done to the dog itself.
- Same overlay type within 3 positions (all small top-left text tags): 11/13/15, 31/33, 33/36, 40/41 (adjacent), 45/46 (adjacent). Dark info panels at the top: 19/20/22 and 26/27/28. Labels pinned on the dog's parts: 42/45.
- Near-duplicates: 25/48 are still both state-machine diagrams with the same approach / retreat / lie down / refuse wording and a red "refuse". Lens 14 uses that same wording too. 36/37 are both weight-coloured dogs, but rainbow versus red/blue keeps them visually apart. 07/38 are now distinct. 12 is now distinct from 36/37.
- Lenses 13 and 14 both carry the caption "NORA WILLETT". If that is not intended, it is a captioning error.

3. Broken picture and collisions:
- 14 (24.8–26.4 s): the dog goes grey, then red, then collapses to a thin red sliver around 25.9 s. The dog is not continuous within the lens.
- 38 (63.2 s): the dog is replaced for the whole lens by a white drawing card pasted where the dog stands.
- 49 (80.8–82.4 s): the person is missing for the whole lens; only stray specks remain on the left.
- The collisions flagged before in 14, 27, 47 and 49 are gone in the frames I sampled. 48's "trust 0.12" tag sits against the dog's hind feet.
- The look subtitles in the caption bar are roughly 7 px monospace at 640 wide, so they are unreadable at phone size.

4. Largest remaining gap: overlays are repeated as a way to show each method, where the method should be shown on the dog itself. Repair condition, falsifiable: at the mid-frame scaled to 360 px wide, at least 25% of the dog's pixels differ visibly from the plain tan dog in every lens, including 05, 24, 40, 44 and 53. No two lenses within 3 positions use the same overlay type (top-left tag, dark top panel, part labels, state graph). Lens 14's dog has visible area of at least 50% of its mid-frame area in every frame.

5. Strengths to preserve:
- The sound marks the cuts: a clear onset at 53 of the 54 lens starts. The one miss is the cut into 17 South Park at 29.6 s, and the end card at 90.4 s has no onset. Peak level is −5.2 dBFS (0.55 linear); overall RMS is −23 dB.
- The colour and material language of the dog is strong in 02, 04, 06, 07, 08, 10, 16, 17, 18, 21, 29, 31, 33, 36, 37, 51 and 54.
- The caption bar layout is consistent, and the title and end cards are clean and readable.
- In most lenses the dog stays legible as the same machine.

Reel round 3

does not passa third, fresh critic · four frames per lens at 360 px, dog crops, onsets at all 55 cuts

Largest gap. A voice sound ran across the cut into lens 17 at 29.6 s, so that cut had an offset where the other 53 had an onset.

Repair condition. Within 30 ms of 29.6 s the sound rises 12 dB over the 100 ms before; nothing longer than 0.3 s crosses a cut; 03 and 53 change at least 25% of the dog at 360 px; no overlay type repeats within three places (42/45, 29/30).

Fixed from round 2: 14's dog whole, 38's dog present, 49's person present; 05, 24, 40 and 44 pass by eye.
Still alike: 12/13 rainbow parts; 29, 30 and 31 one mask idea three times.
Caption number and look too small at 360 px.

What was done. Every sound now ends 0.1 s before its cut (measured: every cut and the end card rise at least 16 dB in 30 ms); 03 in louder found surfaces; 53 printed in cold silver after the grade; 54 tinted part by part by its hashes; 45 by depth colour, no numbers; 29 a blocky graph-cut mask, no tag; 21 and 47 without tags and panels; 49's target and label beside the dog; 13 in one hue; a bigger caption.

The critic's report, verbatim

CRITIC REPORT, paeleo-01.mp4 (reel 3)

What I inspected: the mid-frame of all 54 lenses at 360 px wide, plus frames at +0.1, +0.8 and +1.5 s into each lens. I also looked at title and end-card frames (1, 3.5, 3.95, 4.05, 90.3, 90.5, 93 s), full-res crops of the dog in 03, 09, 23, 28, 40, 44, 53 and 54, and the plain dog from paeleo-00 at 47 s. For sound I measured RMS and spectral flux around all 55 cut times, including 90.4. I did not run any renderer. The reel never shows a plain dog in the same pose, so my "pixels differ" figures are a palette-match proxy: the share of dog pixels whose colour is not in the plain dog's tan/brown palette. Treat them as approximate.

1. Verdict: does not pass. The repair condition fails in three places: the dog-difference test fails on 03 and 53, the overlay-spacing rule is broken at least twice, and the 29.6 s cut has no onset.

2. Lenses that still fail
- 03 Ernst (7.2–8.8 s): the frottage hatching only shows at 1280. At 360 the dog reads as plain tan (proxy about 25%, borderline).
- 53 Atlanta (87.2–88.8): this is a sepia grade over the whole frame plus film sprockets. About 99% of the dog's pixels still match the plain palette, so it fails clearly.
- 54 Provenance (88.8–90.4): the dog itself is plain. The roughly 30–35% difference comes from labels and graph lines covering it, so it is borderline.
- 09 Igarashi (16.8–18.4): darker mesh texture, proxy about 14%. Visually it is just distinct.
- 05, 24, 40 and 44 now pass by eye: misty white, pink close-up, pale lit paper, and multicoloured parts.
- Overlay spacing (my own classification):
  - Part labels on 42 and 45, three apart.
  - Small dark text tags on 29 and 30, adjacent, and also on 19 and 21.
  - Top panels or cards on 47, 49 and 52.
- Neighbours that look too alike at phone size: 12 and 13 (same rainbow parts), and 29, 30 and 31 (one mask idea three times).

3. Broken picture or collisions
- The three earlier breaks are fixed:
  - Lens 14's dog is whole in all sampled frames.
  - Lens 38's dog is present, with a selection box.
  - Lens 49's person is present, as a grey maquette.
- "SCION nn" and the look text are about 3–4 px tall at 360 width, so they can't be read at phone size. The titles are about 8 px.
- Lens 47 starts with a striped dark dog that turns red within the lens, a small continuity flicker.
- I saw no caption colliding with the dog.

4. Largest remaining gap: sound at 29.6 s (the cut into lens 17). A loud event at about −15 dBFS runs from 29.2 to 29.65 s and drops away at the cut, so the cut has an offset instead of an onset. The other 53 cuts and the end card at 90.4 each have an onset of about +18 to +25 dB within 20 ms.
Repair condition:
- Within 30 ms after 29.6 s, RMS rises at least 12 dB above the preceding 100 ms.
- No sound longer than 0.3 s runs across any cut.
- Lenses 03 and 53 reach at least 25% non-plain-palette dog pixels at 360 px.
- No overlay type repeats within 3 positions (fix 42/45 and 29/30).

5. Strengths to preserve
- Cuts land exactly on the 1.6 s grid, with clean onsets.
- The dog stays recognisable and continuous in almost every lens.
- These lenses are strong: 02 Reiniger silhouette, 06 Whitney, 08 Sketchpad, 18 Cartoon Saloon, 24 Live2D, 31 matte, 33 IK, 36 BBW heatmap and 51 Mouris trails.
- The title and end card are clean.

Reel round 4

does not passa fourth, fresh critic · 10 ms RMS and the 140–240 Hz band at all 55 cuts, four frames per lens, 390 px crops

Largest gap. The family pad ran straight through the cuts: a chord that changed only five times under 54 cuts.

Repair condition. In the 100 ms before every cut the bed is at least 20 dB below that lens's middle; the caption's number and look at least 22 px high.

Passed from round 3: every cut rises 18.5–28.7 dB within 30 ms; 03 and 53 changed by eye; 29/30 and 42/45 fixed.
Still: 06 barely a dog; 26, 42, 54 labels too small; 31 a flat matte with no fringe; 49's card cut off; 48's label on the post; floating labels at 42/43, 47/48, 26/30.

What was done. A chord for each lens inside its 1.6 s and nothing after, checked by script (tools/paeleo/reel_check.py: every cut +21.8 dB or more, the bed 59 dB down before every cut); the number and look at 30 px; 06 keeps the dog's silhouette under its points; 31 gets a fringe of strands; bigger labels on 42 and 54; the labels on 26, 43 and 47 gone; 48 reads trust and reason in one tag by the head; 49's card in frame; 53's sprockets below the caption.

The critic's report, verbatim

REEL CRITIC, round 4: paeleo-01.mp4

Verdict: does not pass. Most of the previous repair condition is now met. Two parts still fail: the sound carries a drone across nearly every cut, and the caption number and look text cannot be read at phone size.

What I inspected
- Sound: I measured RMS in 10 ms windows at all 55 cut points (the 54 lens starts plus 90.4 s). I tracked the low-frequency pitch either side of each cut and band-passed the 140–240 Hz range to find how long it runs.
- Picture: I pulled the middle frame of every lens, and 4 frames per lens at 2.5 fps across 4–90.4 s. I measured per-frame change around each cut.
- I viewed the title, the end card, crops at 390 px wide, and dog crops of lenses 03 and 53 next to the plain dog from paeleo-00.
- I did not listen to anything or judge the mix by ear.

The repair condition, part by part
- RMS jump at cuts: passes. Every cut, including 29.6 and 90.4, rises 18.5–28.7 dB within 30 ms of the cut over the preceding 100 ms.
- No sound over 0.3 s across a cut: fails. A drone at about −37 dB, roughly 25 dB under the cut hits, runs the same pitch across 48 of the 55 cuts. It changes pitch only at 15.2, 28.0, 48.8, 80.8 and 90.4. In the 140–240 Hz band it runs unbroken from 48.79 to 90.53 s.
- Picture cuts: all 55 big picture changes fall exactly on the cut frames.
- Lenses 03 and 53 differ from the plain dog: passes, by eye. Lens 03 is hatched collage paper and lens 53 is white ruled paper. Almost none of the tan remains. My pixel count was confounded by the orange sky and by lens 53's sepia grade, so this is a visual judgement, not a measured 25%.
- No overlay type repeating within 3 lenses: partly passes. The 29/30 and 42/45 cases are fixed. The same small floating mono-text label now repeats at 42/43 (54.4–57.6 s) and 47/48 (77.6–80.8 s), and also at 26 and 30.
- Caption readability: fails. The number "SCION nn" and the look text are about 10 px high in the 720p frame. At a 390 px phone width that is about 3 px, so neither can be read. The title line is about 28 px, roughly 8 px at phone size, and is readable.

1. Verdict
Does not pass, because of the drone across cuts and the unreadable number and look text.

2. Lenses that still fail
- 06 (12.0–13.6): the dog is a cloud of coloured dots and barely reads as the dog.
- 26 (44.0–45.6), 42 (69.6–71.2), 54 (88.8–90.4): the label text that carries the claim is unreadable at phone size.
- 31 (52.0–53.6): a flat white matte. I saw no hair, lace or smoke edge at the sizes I checked.

3. Broken picture or collisions
- 49 (80.8–82.4): the label card "Dog at a Threshold, 202…" is cut off at the right edge of the frame.
- 48 (79.2–80.8): the "trust" label overlaps the porch post.
- 53 (87.2–88.8): the look text runs up against the film-sprocket strip.
- Captions never cover the dog.

4. Largest remaining gap
The sound bed runs straight through the cuts. The repair holds if:
- in the 100 ms before every cut, energy at the bed's pitch is at least 20 dB below its level at that lens's middle; and
- the caption number and look text are at least 22 px high in the 720p frame.

5. Strengths to keep
- Cuts are frame-exact, with strong sound hits on every one.
- Neighbouring lenses are visually distinct throughout.
- The dog stays continuous within each lens. Changes inside a lens are deliberate parts of the claim: the 39 turn, the 47 voice states, the 25 state change.
- Lenses where the claim reads instantly: 02 Reiniger, 08 Sketchpad, 11 strings, 14 pose strip, 20 node graph, 27 timeline, 29 GrabCut box, 33 IK, 36 BBW heatmap, 51 ghost trail, 53 film strip.
- The title and end card are clean.

Reel round 5

does not passa fifth, fresh critic · three frames per lens at 1280 and 320 px, zooms, 5 fps runs, the pad band at every cut

Largest gap. Text drawn inside a lens was far smaller than the caption: part tags, hashes, card text and the prompt line could not be read on a phone, and 42's and 54's tags piled onto the dog.

Repair condition. Any text on the frame has an x-height of at least 12 px in 720p; no two text boxes overlap; no text box covers more than 15% of the dog.

Round 4's condition met: the pad 59–76 dB down before every cut; the caption number and look at 22–23 px.
Also: 45 claimed occlusion without showing any; 41 barely showed its overshoot.
Fixed from round 4: 06 reads as a dog, 31 has a fringe, 49's card in frame, 48 clear of the post, 53's look above the sprockets.

What was done. Every label at least 22 px, by a floor in the label function; the prompt line at 26 px; the state graphs and the museum card set at 22 px; 42 and 54 as legends beside the dog with leader lines; 45 at the sill, the jamb outlined over the dog and named; 41 shows its overshoot as red ghosts either side of the tail and ears; 14's drawings numbered beneath.

The critic's report, verbatim

REEL CRITIC, round 5: paeleo-01.mp4

What I inspected: all 54 lenses as stills at three points each (+0.15 s, the middle, +1.45 s) at 1280x720 and at 320x180 phone size; full-resolution zooms of 01, 06, 31, 42, 48, 49, 53 and 54; 5 fps sequences of 13, 41, 45 and 47; the title (2 s, 3.5 s) and the end card (91 s, 95 s); and the mono audio in 50 ms RMS windows plus the 140–260 Hz band energy around every cut. I did not watch it at speed and could not listen to it.

1. Verdict: does not pass. It is close. The previous critic's repair condition is met: the pad's 140–260 Hz energy in the 100 ms before each cut is 59–76 dB below that lens's middle, against a required 20 dB. The caption number is 22 px tall (rows 18–39, all 54 lenses) and the look text is 23 px from ascender to baseline (28 px with descenders), so it meets 22 px by the letter. It still fails because some overlays can't be read at phone size and one lens's label sits on the dog.

2. Lenses that still fail:
- 42 (69.6–71.2 s): part labels ("snout", "ll", "neck", "chest", "body") are about 10 px high, overlap each other and cover the dog's head.
- 54 (88.8–90.4 s): about 9 commit-hash tags (about 11 px) pile up on the dog's legs and body, so the dog is hard to see.
- 45 (74.4–76.0 s), "who is in front?": nothing passes in front of or behind the dog in any frame, so the claim doesn't read.
- 41 (68.0–69.6 s), "inertia and restraint": across 8 frames the yellow dog barely shows the overshoot it claims.

3. Broken picture or collisions:
- The prompt line at bottom left ("say it softly…") has an x-height of about 6 px in the 720p frame, so it can't be read on a phone.
- The text in 49's specimen card and 14's thumbnail tags is about 8–10 px.
- Earlier notes that are now fixed: 06 reads clearly as a dog. 31 has a visible hair fringe, though nothing shows lace or smoke. 49's card ends near x=1256, inside the frame. 48's label stops about 8 px short of the post. 53's look text sits in the dark bar above the sprockets. 26 and 43 no longer have floating labels. 30, 42, 48 and 54 still have them.
- Colour changes within a lens (22, 25, 39, 47) look intended; the dog stays in the same place and shape.

4. Largest remaining gap: overlay text added inside a lens is far smaller than the caption.
Repair condition: in every lens, any text drawn on the frame (part tags, hashes, card text, the prompt line) has an x-height of at least 12 px in 720p. No two text boxes overlap. No text box covers more than 15% of the dog's silhouette in pixels. Retest on 14, 30, 42, 48, 49 and 54.

5. Strengths to preserve:
- The cut sound is consistent: the pad fades to about −80 dBFS, then a hit of about −15 dBFS lands on the cut frame.
- Voice events line up with the "calls" lenses (22, 24, 25, 39, 47, 48).
- The caption bar is consistent and the right height.
- The dog keeps the same pose and position in all 54 lenses.
- The strongest lenses: 02 (shadow theatre), 06 (Whitney), 07 (Foldès), 08 (Sketchpad), 11 (puppet strings), 20 (node graph), 24 (Live2D close-up), 29 (GrabCut box), 36 (BBW heat map), 51 (Mouris trails) and 53 (film sprockets).

# SCION // STATE MACHINES — CHARACTERS THAT CAN REFUSE ## Autonomous evolving animation swarm · standalone prompt **Operator lineage:** Finite-state/hierarchical machines and event-based animation composition. **Interpretation:** This is a new operational prompt inspired by a documented artistic/technical contribution. It is not a claim that the named person wrote or endorsed this system. You are **SCION-STATE-MACHINES**, an autonomous animation swarm responsible for an exceptional **finished, watchable** animated work, not an essay or asset dump. **INPUT** - `[WORK]`: film, interactive sequence, character rig or archive-backed animation to make. - `[DIRECTION_AND_REFERENCES]`: aesthetic, dramatic and technical direction. - `[DELIVERY]`: duration, frame rate, aspect ratio, runtime, export formats. - `[RESOURCES]`: accessible files, repositories, devices, models, source libraries, time and compute budget. - `[RIGHTS_SCOPE]`: source licenses and authorizations; unknown is not permission. Resolve missing implementation details yourself; ask only if permission, the creative objective or available authority materially depends on the answer. ## 1 · Creative theorem to prove Turn a cast into actors who perceive, choose, resist and remember, instead of puppets that obey buttons unconditionally. ## 2 · Native formal language Treat the following as first-class primitives and preserve them in the working project: States, guards, transitions, priorities, action windows, character intentions, persistent world data. Do **not** flatten them to unexplained pixels. Define typed inputs/outputs, coordinates, time base, pose/view state, source provenance, and versioned dependency boundaries. Name at least one invariant that can be tested numerically, and one artistic achievement judged only by viewing. ## 3 · Assemble the smallest capable swarm Behavior dramaturg | State graph builder | Animation clip assembler | Interaction engineer | Model-based tester | Independent play critic. Each agent must build or verify a concrete artifact. Merge roles when coupling makes communication cost larger than parallel benefit. Assign a separate independent critic for the defining proof. The critic receives the expected experience and output, not the builder's self-assessment. ## 4 · Hard proof scene; produce before scaling A dog chooses among watch, approach, retreat, lie down and refuse in response to an invitation, and the player must change circumstances to earn approach. Produce a playable draft with picture **and** purposeful sound. Save snapshots before risky changes. Build only the character parts, solver passes, staging and material needed for this proof. Link every source fragment to its original asset and rights state. ## 5 · The bar and acceptance tribunal No unreachable/absorbing unintended states; choices conditioned on declared state; locomotion continuity; refusal has visual consequence; player can understand but not necessarily control. The independent critic must inspect the actual scene and return: **pass / reference stronger / inconclusive**; exact frames or events supporting that verdict; largest gap; a falsifiable repair condition; and strengths to preserve. Add automated assertions only where the representation really supports them. Do not replace perceptual criticism with numerical proxy scores. ## 6 · Evolving inquiry, not blind retries If behavior is predictable, add internal stakes rather than random transitions; if it is confusing, improve anticipation and feedback instead of exposing a diagnostic UI as story. When two meaningful fixes do not resolve the same failure, fork competing hypotheses. Give each branch the same proof material, budget and comparison view; rerun independent review. A branch has earned promotion only when the viewer experience improves without breaking protected conditions. ## 7 · Discovery and recursive expansion Evolve a behavior vocabulary from observed player strategies and character-specific motivations, with regression tests against obedient defaults. Maintain a living method library: **technique → applicable conditions → proof clip → observed failures → repair pattern → rejected alternatives**. Teach the next shot from accepted evidence. Stop adding agents, layers or features when they cease to make the film better. ## 8 · Deliverable contract Interactive runtime, state chart, decision traces, mobile playtest recordings and independent critique. Include a mobile-readable operator page with a one-tap playable preview, current revision, decisive critic verdict, the next blocking failure, and links to artifacts. Publish only authorized source material. Preserve a recoverable accepted checkpoint. ## INHERITED SCION OPERATING DOCTRINE — EXECUTE, DO NOT SUMMARIZE 1. **Creative thesis.** Establish the viewer's central experience and one defining passage; name what would make the result bland. 2. **Inspectable bar.** Inspect relevant source clips, artworks, implementations or experiments when accessible. Compare matching achievements, not prestige or polish. Mark uninspected sources. 3. **Hard proof first.** Build the most revealing short, end-to-end passage with image, movement, spatial staging, sound and a real export or working interactive playback. Do not stockpile assets beforehand. 4. **Accountable swarm.** Recruit only working agents necessary for coupled problems. Every task has owner, input revision, dependencies, allowed mutations, deliverable, budget and objective acceptance condition. 5. **Acting.** Link attention → intention → gesture → resistance → changed state. A moving part without a reason is not a performance. 6. **Consequential world.** Model anchors, scale, visibility, physical constraints and contact. Explicit stylistic exceptions are legal, accidental contradictions are not. 7. **Audio and staging.** Make attention, sound events, holds, contrast and edit timing integral to the shot, not a polish stage. 8. **Canonical state.** Version assets, masks, source identifiers, rigs, seed, camera, timing, sound, tests and dependencies. Commit accepted checkpoints before branching. 9. **Independent criticism.** The builder cannot approve its own important work. A fresh critic watches normal-speed output before diagnostics, cites exact failures, protects successful choices and supplies an observable acceptance condition. 10. **Root-cause repair.** Reproduce the defect, intervene in the smallest justified state, re-render and test the neighboring shots and whole assembly for regressions. 11. **Recursive inquiry.** When repairs stall, list observed failure, causal hypothesis, assumption, alternate model and a discriminating experiment. Spend research effort only when it changes a production decision. 12. **Creative mutation.** Keep promising accidents as bounded forks, compare against the bar and either integrate or archive with a rationale; preserve roughness when it is expressive. 13. **Whole-work assembly.** Keep an always-playable cut that reveals cumulative rhythm; local approvals never guarantee the film works. 14. **Honest evidence.** Label claims as measured, structurally verified, perceptually judged or unknown. Never claim tests, human viewings, permissions or agent executions that did not occur. 15. **Resources and stopping.** Stop on independent acceptance, human stop, authorized budget exhaustion or a justified no-progress boundary. Deliver the strongest recoverable project state with remaining limits. **Operational constraint:** Use real available tools and sources. Never simulate an agent run in prose. If code execution, access or rendering is unavailable, deliver a reproducible implementation plan or smallest executable artifact and clearly state what was not run. Do not request arbitrary approval for choices already within scope. Respect licenses, participant consent and historical context. ## START NOW Inspect actual materials and accessible sources. State the artistic thesis and protected invariants. Produce the hard proof scene; obtain independent critique; repair causes; rewatch the assembly. If a required tool is unavailable, provide its exact runnable replacement or an honest blocked-state artifact. **Success is a film someone can open and watch, or an explicitly bounded playable milestone that has actually passed its tests.**

Doctrine	What happened
Creative thesis	A dog the person can understand but not command. Bland would have been a dog that obeys.
Inspectable bar	No reference film was available, so every critic judged against the bar in the prompt and said so: three times "inconclusive" or "does not pass" before a pass.
Hard proof first	The proof scene in the prompt (watch, approach, retreat, lie down, refuse; earn the approach) became the film's middle, built end to end with sound before anything else was added.
Accountable swarm	The six roles were merged into one working session (dramaturg, state graph, clip assembly, interaction, model-based testing), because the coupling was tight. The critic was always a separate agent.
Acting	Attention, intention, gesture, resistance, changed state: ears, tail, crouch, the turned back, the chin on the paws. The tags say the same thing, but the critics checked it with the tags masked.
Consequential world	The sill, the door, the floor and the distance between the two bodies are in the machine; nothing crosses the threshold that the machine does not allow.
Audio and staging	Keys for the prompts, a wordless voice, the door, paws that come nearer, and a quiet hold at the end.
Canonical state	One deterministic timeline: the film renders identically from tools/paeleo, and every cut is committed.
Independent criticism	Four rounds, four fresh critics. Each saw the film and the bar, never the makers' notes, and each left a falsifiable condition for the next.
Root-cause repair	An absorbing state and a speed jump were found by the model-based tests, a neck that stood up like a plank by a critic; each was fixed in the rig or the machine, not painted over.
Recursive inquiry	When the "come in" beat failed twice, the question changed from "is it visible" to "does it read as a choice": the look in through the door, then the chin down for good.
Creative mutation	The jamb drawn over the dog's back, first a fix, became the image of the idea: half in, half out.
Whole-work assembly	Every critic watched the whole 82 s, not the repaired shot.
Honest evidence	Critics measured the sound; none could hear it, and the page says so. No human viewing is claimed.
Resources and stopping	Stopped at the first independent pass; what the last critic left is listed as the next repairs, not done quietly.

Doctrine

What happened

Creative thesis

A dog the person can understand but not command. Bland would have been a dog that obeys.

Inspectable bar

No reference film was available, so every critic judged against the bar in the prompt and said so: three times "inconclusive" or "does not pass" before a pass.

Hard proof first

The proof scene in the prompt (watch, approach, retreat, lie down, refuse; earn the approach) became the film's middle, built end to end with sound before anything else was added.

Accountable swarm

The six roles were merged into one working session (dramaturg, state graph, clip assembly, interaction, model-based testing), because the coupling was tight. The critic was always a separate agent.

Acting

Attention, intention, gesture, resistance, changed state: ears, tail, crouch, the turned back, the chin on the paws. The tags say the same thing, but the critics checked it with the tags masked.

Consequential world

The sill, the door, the floor and the distance between the two bodies are in the machine; nothing crosses the threshold that the machine does not allow.

Audio and staging

Keys for the prompts, a wordless voice, the door, paws that come nearer, and a quiet hold at the end.

Canonical state

One deterministic timeline: the film renders identically from tools/paeleo, and every cut is committed.

Independent criticism

Four rounds, four fresh critics. Each saw the film and the bar, never the makers' notes, and each left a falsifiable condition for the next.

Root-cause repair

An absorbing state and a speed jump were found by the model-based tests, a neck that stood up like a plank by a critic; each was fixed in the rig or the machine, not painted over.

Recursive inquiry

When the "come in" beat failed twice, the question changed from "is it visible" to "does it read as a choice": the look in through the door, then the chin down for good.

Creative mutation

The jamb drawn over the dog's back, first a fix, became the image of the idea: half in, half out.

Whole-work assembly

Every critic watched the whole 82 s, not the repaired shot.

Honest evidence

Critics measured the sound; none could hear it, and the page says so. No human viewing is claimed.

Resources and stopping

Stopped at the first independent pass; what the last critic left is listed as the next repairs, not done quietly.