Can a description make a character that chooses, so that the person who prompts it can understand it but not command it?
Nobody animates the dog. The prompts change the world (posture, place, door, voice, hand); a five-state machine reads that world through what it perceives, trusts and remembers, and the paper puppet acts out whatever it chose.
You are the operator. You cannot move the dog. You can only change the circumstances, the way the prompts do in the film. It is the same dog: the same machine and the same paper, running in your browser.
Five states. Guards are checked in priority order every frame; each state holds for a minimum window so the dog never flickers between choices. Danger (retreat, refuse) may cut a window short; nothing else can.
| Guard | Goes to | When |
|---|---|---|
| refuse | turns its back | pressure > 0.62 from repeated calls, or pressed (threat > 0.55) when already 3.6 m away or at the steps |
| refuse → watch | turns back | held 3.5 s, pressure < 0.2, invitation > 0.5 and threat < 0.25 |
| retreat | backs away, facing the person | threat > 0.55 within 3.6 m (may cut any window short) |
| lie down at the sill | lies through the doorway | arrives at the sill, or trusted and unthreatened there |
| lie down → watch | gets up | threat > 0.45 while lying |
| approach | comes closer, slower when unsure | trust > 0.6, threat < 0.3, the threshold clear (the person sitting, or not in the doorway) |
| approach → watch | stops to look | threat ≥ 0.3 or trust < 0.5 on the way |
| lie down where it is | settles | nothing asked of it (silent, no hand, no threat) for 6 s |
| watch | watches | otherwise |
Run with node tools/paeleo/test_machine.js.
Every choice the dog made in the film, with the guard that made it. Film time is machine time plus 18 s.
| Film time | Prompt or choice | Why |
|---|---|---|
| 21.0 s | › the person stands in the doorway and calls: Come here. | prompt takes effect: voice = call |
| 21.0 s | watch → retreat | threat 0.74 > 0.55 at 3.3 m: backs away (trust 0.17, at 3.00 m) |
| 22.4 s | retreat → watch | watches (trust 0.12, at 3.95 m) |
| 24.8 s | › call again, and reach for it. | prompt takes effect: voice = call, hand = reach |
| 24.8 s | watch → refuse | threat 1.00 and already 4.3 m away: turns its back (trust 0.10, at 4.06 m) |
| 31.6 s | › step back. sit down beside the door. say nothing. | prompt takes effect: place = beside, posture = sit |
| 33.5 s | › open the door. | prompt takes effect: door = open |
| 33.5 s | refuse → watch | pressure 0.00 gone, invitation 0.60: turns back to look (trust 0.12, at 4.06 m) |
| 37.6 s | watch → lie down | nothing asked of it for a while: settles where it is (trust 0.31, at 4.06 m) |
| 39.5 s | › offer a hand, low. | prompt takes effect: hand = low |
| 42.6 s | › say it softly: it's all right. | prompt takes effect: voice = soft |
| 43.9 s | lie down → approach | trust 0.60 > 0.6, threat 0.00, threshold clear: gets up and comes closer (trust 0.60, at 4.06 m) |
| 53.1 s | approach → lie down | reached the sill: lies down there, half in and half out (trust 0.83, at 0.59 m) |
| 56.0 s | › come in. | prompt takes effect: voice = soft |
| 56.0 s | lie down (stays) | hears the voice: head and ears come up, the tail goes, then the chin goes down on its paws. No guard fires: nothing it perceives has changed enough to move it. |
Every cut, in order.
The whole film end to end: the lineage, the title, seven prompts, the three scales, the lockup. The dog makes seven choices on its own.
Repairs for every cause the first critic found, then a fresh critic.
Repairs for the second critic, then a third.
Repairs for the third critic, then a fourth.
An independent critic watched each cut without the makers' notes and returned a verdict, the moments behind it, the largest gap and a condition a re-check could test.
Largest gap. The thesis beat is told by text rather than acted: told to come in, the dog does nothing you can read as a choice.
Repair condition. In the 2 s after "come in." resolves: (a) the dog visibly reacts, ears or head toward the doorway or the tail moves; (b) then settles back clearly, head onto paws or turned back out; (c) on the last frame part of the body is inside the lit opening and part outside. Legible at 320 px wide with the tags masked.
Keep. The cut paper and dusk palette; the P/E/O bar and tags that update; the close-up refusal at 25–28 s; the person sitting; the continuous approach with paws in sync; the door creak; the nesting scales and the lockup.
What was done. Every cause repaired (see What broke and what changed); round 2 went to a different critic.
CRASH LABPAELEO-00 v1 critic report (round 1)
What I inspected: frames at 2 fps across all 82 s (contact sheets), 8 fps tiles at 4.6–6.1, 21–23, 32.6–33.8, 37.2–38.8 and 52.3–53.5 s, a 24 fps crop at 33.3–33.8 s, and full-res stills. For sound I measured RMS, peak and spectral centroid every 0.25 s, rendered a full spectrogram, zoomed in on six moments and checked a 5 ms envelope. I cannot hear it, so I'm not claiming the voices are intelligible as speech.
1. Verdict: inconclusive. No reference film was supplied to compare against, so "reference stronger" can't be judged. Measured against the bar, it does not pass: the climax refusal has no acting, and the dog visibly pops at 33.5 s.
2. Moments behind the verdict
- 54.1–59.6 s (about frames 1298–1430): after "come in." the dog's pose is the same in every 0.5 s sample. No ear, head, tail or glance toward the doorway changes. Only the camera drifts. The refusal is carried by the narration typed at 58.0–60 s ("the dog stays on the sill. the description let it choose."). The dog lies in profile, parallel to the lit door, so "half in, half out" isn't staged at all.
- 33.50 s (frame about 804): the dog goes from lying curled and facing away (large) to an edge-on sliver in one frame, then stands up much smaller in the same framing. That is a pose and scale snap.
- 37.5–38.0 s: the dog's response to "open the door" (lying down, tag says "lying down") only lands while the next prompt "offer a hand" is already being typed and the operation light is on. That muddies which change caused it.
3. Single largest gap
The thesis beat, "the viewer can understand the dog but not control it", is told by text rather than acted. When asked to come in, the dog doesn't do anything you can read as a choice.
4. Repair condition (falsifiable)
In the 2 s after the "come in." prompt resolves, a re-check at 8 fps must show:
- (a) the dog first reacts visibly: ears or head turn toward the doorway or interior, or the tail moves;
- (b) then it settles back clearly: head lowers onto paws, or it turns its head back out;
- (c) on the final frame of the scene, part of the body (front paws or chest) is past the door-frame line inside the lit opening and part is outside.
All of this has to be legible in a 320 px-wide still with the tags masked.
5. Other problems, ranked
1. 33.50 s (frame about 804): the lying-to-standing snap and scale pop. Every frame-to-frame pose change needs to be continuous, at least 6 frames.
2. 37.5 s: make the reaction to the door read before the next prompt starts typing (an approach or a lie-down inside 34–36 s).
3. 28.0–33.4 s: in the wide shot, the "turned its back" dog is a small curled lump about 70 px wide at 1280. It doesn't read as turned away at phone size; only the close-up at 25–28 s does.
4. 5.0–5.9 s: the IMAGETEXT caption doesn't lift off. The growing picture covers it and it vanishes, then it reappears at 5.9 s as a second copy beside a new tag. That's two copies of the same text over the door frame.
5. 18.0 s and 6.0 s: crossfades overlap text. The title is ghosted over the tags, and "ENVIRONMENT TEXT" is doubled.
6. 64.5–71.0 s and 77.5–81.5 s: the subtitles in the three-scales diagram, the cycle labels and the credits are tiny monospace, unreadable at phone size. All tag text is also around 11 px at 720p.
7. 21.0–22.4 s: "backing away" is a slow drift of roughly 15% of frame width with the head low. It's legible, but the ears and tail don't help.
8. Sound: the voices are short synthetic harmonic glides (21.5, 25.0, 41.5–43, about 54.5 s). The soft voice at 41.5 is noticeably lower and quieter, which is good. "Come in" at 54.5 s is mostly a broadband burst with little harmonic content, so the climax line is the weakest sound cue. Paw-like clustered transients are present at 44–51.5 s.
6. Strengths to keep
- The cut-paper look and the dusk palette.
- The P/E/O indicator bar and tags that update with each prompt.
- The close-up at 25–28 s (dog crouched, head down, facing away). It's the best piece of acting in the film.
- The person sitting down at 31–32 s.
- A continuous approach walk with matching paw sounds (43.5–53 s) and a smooth lie-down onto the sill (53.0–53.4 s).
- A door creak synced to the door opening (33.5–35 s), typing sounds synced to each prompt, a cricket ambience bed, and music only under the close.
- The three-scales nesting animation and a clean PAELEO lockup with a readable tagline (72–77 s).
Largest gap. The refusal of "come in" is a near non-event: a small head lift outward, head down, then up again. It never looks into the house, so refusal cannot be told from not hearing.
Repair condition. Within 1 s of the prompt: the head turns at least 30° toward the lit interior and holds 6 frames, or the tail sweeps twice; within 1.5 s more it turns back out and lowers onto the paws; it does not rise again before the scene ends; in a 320 px still of the last frame the dog crosses a jamb with at least a quarter of its length on each side.
Keep. The caption lift-off as one travelling object; the refusal close-up; the continuous approach and lie-down; the clean head-lower at 57.3 s; the nested scales card; the title card.
What was done. The dog now turns its head in through the door with its tail sweeping, then puts its chin down and keeps it down; the sill moved so its back third lies outside the jamb; the reframes became moves; the sit squats first; the turn waits until the dog is up and takes longer; a pad and tag sounds under the lineage; tail thumps only when the tail sweeps; bigger type.
PLAY CRITIC REPORT: paeleo-00-v2.mp4
What I inspected: a 1 fps contact sheet of the whole film; frame-accurate grabs at 8 fps for 3.5–7, 13.75–18.25 (at 4 fps), 31–36 and 44–64, and at 24 fps for 31.5–32.1, 32.9–33.4, 33.4–34.4, 55.8–56.3, 57.25–57.75 and 59.1–59.6. I also ran per-frame motion and pan-shift measurements over 8–64 s, took per-0.25 s RMS for the whole film, and rendered a log spectrogram. I had no reference cut, so I cannot compare against one.
1. VERDICT: Inconclusive, because there was no reference to compare against. Against the bar on its own terms it does not pass. The climax repair condition is only partly met, and there are visible pops and jump-reframes in the middle act.
2. REPAIR CONDITION TEST ("come in." finishes typing at about 55.9; the speech bubble shows 56.5–57.5)
- (a) Partly met. At 56.00–56.17 the head lifts slightly and the ears prick, over about 4 frames. But the head stays facing out toward the person and never turns toward the interior. The tail does not move: frame difference in the tail region is under 0.8 from 56 to 60.
- (b) Met, then undone. At 57.29–57.75 the head lowers smoothly onto the paws over about 11 frames and reads clearly. But at 59.2–59.9 the head comes back up to an alert, outward pose and holds there to the fade at 63.0–64.0. The scene ends on alert watching, not on a settled dog.
- (c) Marginal. In the 62.9 still at 320 px the body sits inside the lit opening. The forepaws are at the left jamb by the person's feet, and a thin tail crosses past the right jamb onto the siding. It reads as "dog lying in the doorway", but in/out is not legible. The dog is low-contrast brown on yellow, and the "dog · lying on the sill" tag and its leader line sit right above it.
3. LARGEST GAP: The refusal of "come in" is acted as a near non-event. The dog gives a tiny outward head lift, puts its head down, then lifts it again. It never looks into the house and never chooses, so without the tag and the closing line ("the dog stays on the sill. the description let it choose."), the viewer cannot tell refusal from not hearing.
4. REPAIR CONDITION FOR THAT GAP (check at 8 fps, tags masked):
- Within 1.0 s of the prompt resolving, the head turns at least 30° toward the lit interior and holds at least 6 frames, or the tail makes at least 2 visible sweeps.
- Within the next 1.5 s, the head turns back out and lowers onto the paws.
- From then to the last frame of the scene, the head does not rise again: motion in the head region stays under 1.0 mean-abs per frame.
- In a 320 px still of the last frame, the dog's silhouette crosses a jamb clearly, with at least 25% of its body length on each side.
5. OTHER PROBLEMS, ranked:
- 32.958 and 37.208: hard reframe jumps of about 60 px within one shot (pan shift ±60 measured in one frame; 37.2 also slightly rescales). They read as jump-cut pops. Also 28.208 is an abrupt cut back to wide.
- 31.83–32.0: the person's sit pops. A squat snaps to a chair-sit hovering in mid-air at 31.875, then slides to the floor by 32.04.
- 6.30–6.45: crossfade shows two copies of the caption at once, the serif "A dog at a door." stacked over the mono "a dog at a door". About 14.5 and 17.5: the title ghosts over half-faded tags. About 24.75: "Come here!" and "Come HERE!" overlap.
- 33.4–34.0: the dog goes from lying to standing and does a paper-flip turn in about 0.6 s. It is continuous, not a pop, but too fast to read as a decision at phone size.
- Sound: near-silence at about −44 dB from 5.5 to 11 s, under the ENVIRONMENT TEXT and WORLDTEXT beats that carry the lineage. A loud low pulse at about 2 Hz (−17 dB) starts at about 54 s, under the intimate climax. The tail barely moves, so the pulse does not match tail thumps, and it pre-laps the closing music.
- Credits text (79–82) and the 10 px tag body text are unreadable at phone size. In the PEO diagram at 70 s, "description" runs into the red box border.
6. STRENGTHS TO PRESERVE:
- The caption lift-off at 4.5–6.2 reads as one object travelling: the caption leaves the page while the camera pushes into the picture.
- The 55-frame close-up beats at 25–28 (dog backing off, then turning its back) clearly show refusal and its consequence.
- The approach at 44–54.5 is continuous walking locomotion ending in a smooth lie-down on the sill.
- The head-lower at 57.3–57.7 is clean.
- The nested AEL/PAELEO/PEO card (66–71) is legible and well layered.
- The title card is clean.
Largest gap. The last image of the dog: chin on the floor, but the neck stands up behind it like a plank with the collar on it, for 5 s. And only a fifth of the dog lies outside the jamb, so it reads as lying in the doorway, not across the sill.
Repair condition. Every 8 fps frame from 58 to 63 s: the neck joins head to shoulders with no gap and the collar below the shoulder line; at 320 px, a quarter of the dog on each side of the jamb's outer edge; the dog's tag clear of the threshold tag.
Keep. The lineage reading at 1 fps; refusal and its consequence; the approach staging; the ear-prick and settle at the climax; legible credits.
What was done. Chin down now brings the neck base down to the chest; the sill moved so a third of the dog lies outside the jamb, which is drawn over it; the dog's tag sits below the threshold tag; the sit is slower and in two phases; faster title fades; bigger diagram sublines; a flick and a sigh answer "come in".
CRITIC REPORT, paeleo-00-v3.mp4 (round 3) What I inspected: a 1 fps contact sheet of the whole film; 8 fps frames from 52–64 s and 30–39 s (full frame plus dog crops); frame-difference scans over both windows; stills at 6.3, 6.45, 14.5, 16, 17.5, 24.75, 63.0 (also at 320 px), 69 and 79 s; per-0.25 s RMS; a log spectrogram; envelope periodicity for several spans. I did not mask the tags. I cropped tight on the dog instead. I saw no reference film, so I cannot compare against one. 1. VERDICT: DOES NOT PASS. The climax meets 3 of the 4 parts of the previous critic's repair condition. It fails the jamb-crossing part. The final pose also has a visible rig break, so the last image of the story fails the "no pops / acting carries it" bar. 2. CLIMAX AGAINST THE REPAIR CONDITION (8 fps) - The "come in." prompt resolves at about 55.9 s, and a small speech bubble shows 56.0–57.1. - Head: the dog was lying with its head forward toward the lit interior. It lifts the head about 45–60° up toward the doorway light, ears pricked, and holds 56.25–57.5 (10 frames). This passes in letter. It reads as a lift, not a turn, because the head already faced inward. - Tail: it swings up/back/up about twice between 56.0 and 57.6. Marginal pass. - Head lowers 57.75–58.0, within 1.5 s. Pass. - The head does not rise again through 63.0, and the fade runs 63.0–64. Pass. - Jamb crossing: in the 63.0 still the dog sits almost wholly inside the frame of the door. Only the rump, a hind leg and the tail cross the right jamb, about 20–22% of nose-to-rump length. At 320 px it reads as "lying in the doorway", not "across the sill". FAIL. 3. LARGEST GAP: the final pose (58.0–63.0). The head drops to the floor, but the neck segment stays as an upright plank with the collar on it, standing above and behind the head. The head looks detached and the dog looks folded or broken (compare the clean pose at 57.25 with the one at 60.0). This frame is the film's last word on the dog, and it holds for 5 s. 4. REPAIR CONDITION: in every 8 fps frame from 58.0 to 63.0, the neck joins head to shoulders with no gap and the collar sits below the shoulder top line. In a 320 px still at 63.0, nose-to-rump length has at least 25% on each side of the right jamb's outer edge. Between 53.5 and 63.0 the "dog · lying on the sill" tag does not overlap the "threshold · open" tag. 5. OTHER PROBLEMS, RANKED a. Tag collision through the whole climax, about 53.5–63: the dog tag covers the threshold tag, leaving a truncated "s open". b. Person's sit is a 2-frame drop, standing at 31.625 and seated at 31.875. It reads as a collapse. c. Get-up at 33.375→33.5: the lying, head-tucked pose re-arranges to head-right in one frame, landing on the same frame as the door opening. It then rises in about 3 frames. The turn at 34.125–34.25 is an edge-on paper flip, which is acceptable. d. Title crossfades at 14.5 and 17.5: the semi-transparent "THE THRESHOLD" sits directly over the person and door and muddies both. e. Diagram at 69 s: gray sublines are about 7 px tall at 640 wide, illegible at phone size. PAELEO is also expanded as "Lab / Language for Ekphrastic Operations", which does not match the brief. f. Sound: the 2 Hz low pulse now spans about 44–55 s (envelope peak 2.0 Hz). It overlaps the dog's approach, so it may be footfalls; I cannot confirm intent. 5.5–11 s is no longer near-silent: a pad pulsing at about 1.2 Hz, at -23 to -35 dB. During the climax hold, 60–63 s drops to about -42 dB with only ambience and crickets. The cue that answers "come in." is a small tonal blip at about 59.8 s. g. The previous critic's reframe "jumps" near 33 and 37 s: the frame-differences show a continuous push-in, not jumps. Resolved. Checked and clean: 6.3–6.45 (caption and section header don't collide) and 24.75 (tags clear). 6. STRENGTHS TO PRESERVE - The lineage reads at 1 fps: page caption, then the tag pinned in the scene, then WORLDTEXT with the "may watch, approach, back away, lie down, or refuse" tag. - Refusal has a consequence: the dog backs off at about 21 s, then turns its back and lies down at about 25–30 s. - The approach sequence (door, hand, soft voice) is well staged. - The alert ear-prick at 56.25–57.5 followed by settling back down is the right acting beat. - The credits are legible at 640 px.
Largest gap. Not blocking: in the final sill pose the forelegs stack over the neck and one crosses the collar, so at phone size the front end reads as a knot with an eye rather than a chin on the paws.
Repair condition. The next cut: from 58 to 63 s no foreleg overlaps the neck or collar, the collar is one band, both front paws lie ahead of the chest and under or ahead of the chin, and at 320 px the chin visibly rests on a paw.
Keep. The lineage; refusal and its consequence; the patient approach; the dog half in and half out at the jamb, which states the idea in one image; a look, then a settle, so the film does not obey; the quiet hold; the lockup.
What was done. Accepted. The film on the page is v4. The critic's remaining notes are the next box's first repairs, listed under Next.
PLAY CRITIC, round 4: paeleo-00-v4.mp4
What I inspected: 1 fps and 2 fps contact sheets of the whole film with timestamps. 8 fps crops from 52–58 (climax), 58–63 (the neck test, 40 frames) and 53.5–58 (the tags). Full-frame stills at 60.0 and 62.5, and a 320 px still at 63.0. 6 fps sheets from 27–35 and 45.5–49.5, full 24 fps from 33.3–34.5, and 12 fps from 24–26.5 and across both title crossfades. A 640 px still at 69.0 and a 360 px still at 74.0. Per-0.25 s RMS for the whole track, plus a log spectrogram. In those last three ranges (33.3–34.5, 24–26.5, the crossfades) I looked at sampled frames only; I did not view every frame.
1. VERDICT: PASS. All parts of the previous critic's repair condition and the climax condition hold, and I found no pops in the ranges I sampled.
2. SUPPORTING MOMENTS
- 58.0–63.0, every 8 fps frame: the dog holds the same sill pose. The neck piece runs unbroken from shoulder to head, and the collar sits well below the shoulder top line. At the 4x zoom of 59.5 there is no gap.
- 62.5 and 63.0 (320 px): nose is about 183 px left of the right jamb's outer edge and the rump about 102 px right of it, so 36% is outside. Pass.
- 53.5–63.0: the "threshold" tag (y≈85–175) and the "dog · lying down" tag (y≈225–315) never overlap during the camera push.
- Climax: "come in." finishes typing at about 56.25. The head lifts at about 56.4 toward the left and the lit interior, and holds through about 57.5 (well over 6 frames). It lowers onto the paws at about 57.7, then stays down until the fade at about 62.8. Pass.
- 31.8–32.3: the person's sit goes through crouch frames and is continuous.
- 33.3–34.5: the dog gets up through intermediate frames. It turns by flipping edge-on at about 33.9–34.0, which reads as a cut-paper flip, not a pop.
- 14.0–14.7 and 17.2–17.9: clean crossfades of about 0.5 s, and the tags fade back in after the title.
- 69.0: the AEL / PAELEO / PEO nesting reads at 640 px.
- Sound: there is a gesture at each prompt, with peaks at 21.0–21.5 (call), 24.75–25.25 (reach), 56–57 ("come in"). It drops to about -40 dB during the sill hold (59.5–63.5), and music returns for the close. The quiet hold works.
3. LARGEST REMAINING GAP
The sill pose, which the film rests on for about 6 s, is a jumbled shape at phone size. Both forelegs are stacked over the neck and chest, and one upper leg crosses the collar, so the collar shows as two separate red bars (see 59.5 and 62.5). At 320 px the front end reads as a brown knot with an eye, not clearly "head resting on paws".
4. REPAIR CONDITION
In every 8 fps frame from 58.0 to 63.0:
- no foreleg segment overlaps the neck or collar;
- the collar is one continuous band;
- both front paws lie forward of the chest and under or ahead of the chin.
In a 320 px still at 60.0, the chin visibly rests on a paw, meaning the paw outline touches the lower jaw.
5. OTHER PROBLEMS, RANKED
1. Climax look, 56.4–57.5: the lifted head turns to a near-frontal three-quarter view, so it partly reads as looking at the camera rather than into the room. A clearer profile toward the interior would make "looks in" unambiguous.
2. Diagram at 69: the labels "intent / description / consequence" and the "THE PLACE / THE METHOD" kickers are about 4–5 px tall at phone width. The P→E→O arc also never closes back to P, so the recursion is not shown.
3. Porch ambience, 18–63: a broadband noise floor of about -30 dB reaching up to 16 kHz could read as hiss on phone speakers.
4. Title at 14.5–17.5: the small red line above "THE THRESHOLD" is unreadable at phone size.
5. 25.4–25.6: at the "turning its back" flip, the dog is a thin edge-on sliver for about 2 frames. That is acceptable, but the turn-away depends on the tag to read.
6. STRENGTHS TO PRESERVE
- The lineage from imagetext to environment text to worldtext (0–14) is clear, with smooth caption-to-tag transitions.
- Refusal has visible consequence: the dog backs away at 21 and turns its back and lies facing away at 25–27.
- The approach (40–49) is continuous and patient.
- The dog lies half inside and half outside at the jamb, which states the idea in one image.
- "come in." gets a look, then a settle. The film does not obey.
- The quiet audio hold at 60–63.
- The lockup at 74 reads at 360 px, and the credits are clean.
Accepted at round 4. What the last critic left for the next cut, in order.
SCION 48, State Machines: Characters That Can Refuse, verbatim from the SCION operator. Its proof scene is this film's middle: a dog chooses among watch, approach, retreat, lie down and refuse, and the player has to change circumstances to earn approach.
# 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.**
paeleo/threshold/machine.js the dog: perception, memory, guards, locomotionpaeleo/threshold/stage.js the paper: set, dog, person, tagspaeleo/threshold/try.js the Try it pagetools/paeleo/film.js the score: prompts, camera, text, sound cuestools/paeleo/render.js renders the score in Chromium to H.264tools/paeleo/sound.py the soundtrack from the cuestools/paeleo/test_machine.js the model-based testspaeleo/scion/operator.html the SCION operator, all 54 prompts