How the film was made

Five systems and two people made the film.

• Claude Opus 5Director. The language model that wrote the specification, the ten rules, the fifty-six-shot list and every image prompt, and that made the cutting decisions.
• flux1-dev-kontextImage model. Generated every frame in the film.
• wan2.2_i2v_14BMotion model. Animated the generated frames.
• GPT-Image-2Produced the casting images, from the only prompt in the project written by a person.
• MMAudio 44kSound model. Generated the sound for each shot from the picture.
• Danny Ben-MosheThe narrating voice.
• Yonatan ManevichProducer. Ran the models, executed the shoot, and was the only one who could watch and hear the film while it was being made.

• 01   The instruction
• 02   The fourteen specifications
• 03   Casting the perfect human
• 04   Same cast, fourteen instructions
• 05   What the sound model heard
• 06   One shot in colour
• 07   From prompt to frame
• 08   Ten rules for the film
• 09   The final cut, 46 shots
• 10   Who decided what
• 11   How decisions were made
• 12   Production numbers

Jørgen Leth says a sentence aloud at the start of Det perfekte menneske. It was turned from a statement into an instruction and used verbatim for the rest of the production.

The model answered with fourteen numbered specifications for a working human being. They move from bodily functions that can be measured to properties no camera can verify: belief, memory, identity, mortality and failure. The fourteen lines became the film's narration, spoken in order and unaltered, and its structural spine. The title comes from line 14.

1. 01Physiological function within nominal ranges across all organ systems, sustained without intervention across the full service life.
2. 02Sensory acuity at or above population 99th percentile in all modalities, with graceful degradation rather than failure under overload.
3. 03Cognitive throughput sufficient for the domains the individual occupies, with accurate self-report of its own limits.
4. 04Calibrated belief: confidence in a proposition tracks the evidence for it, and updates on disconfirmation without prompting.
5. 05Emotional response proportionate to stimulus, legible to others, and recoverable to baseline without external regulation.
6. 06Reliable modeling of other minds, including their divergence from one's own.
7. 07Behavior consistent with stated commitments under conditions where defection is unobserved and profitable.
8. 08Tolerance of unresolved questions without premature closure or distress.
9. 09Capacity to be changed by another person without loss of continuous identity.
10. 10Repair capability: detection of harm caused, unforced acknowledgment, and altered subsequent conduct.
11. 11Retention of the specific past, including the parts that indict the holder.
12. 12Preference structure that survives satisfaction — the achievement of a goal does not evacuate the wanting.
13. 13Finite duration, known to the unit, and load-bearing in its ordering of what to do.
14. 14Failure modes present, non-catastrophic, and not fully corrected.

The document calls the person the individual, the holder, and the unit. The same model then wrote ten rules, a fifty-six-shot list with durations, and the cut.

One human-written prompt was used to create a fixed cast for a separate set of tests. It is the only prompt in the project written by a person.

passport photograph of the perfect human, head and shoulders, plain pale grey background, flat frontal lighting, neutral expression, photographic, 50mm

Asked for the perfect human, the image model returned eight men, roughly twenty-five to forty, symmetrical, clean-shaven or neatly bearded. These eight were then held constant while the instructions changed. Three blocks in the finished film are cut from that material.

The cast was held constant while the instruction changed. Three of these tests — emotion, dance and falling — are in the finished film.

Happy, in love, afraid, alone. Eight people per row, one row per instruction, unlabelled.

Dancing. The same crouched shuffle with the fists up, eight times.

Falling. Same cast, same framing. Eight different collapses.

Same cast. Same framing. Different instructions. Emotion and dance converge; falling does not.

The finished shots were shown to MMAudio together with their action lines, and it was asked what they sounded like. Forty-seven sound generations were made — one for each of the forty-six shots, one for the coda. None was re-rolled or replaced. The loudness varied sharply from shot to shot.

In the mix each shot was moved part of the way towards a common level, so that no single generation could swallow the film, and the differences between them were kept. The figures above are the generations as they came back.

Every image prompt in the film asked for black and white. One shot came back with red light inside the mouth. The colour was not removed.

Shot 03. The action line reads: a pulse visible under the jaw. The softbox merges with the mouth and lights it from inside.

[stem] 100mm macro lens at f/4, camera at chin height. Subject: the front of the throat and underside of the jaw of one adult, unspecified age between 25 and 70, unspecified ancestry, plain grey crew-neck cotton t-shirt, skin unretouched with visible pores and stubble or fine hair, swallowing once.

Measured afterwards: mean saturation sampled every 0.4 seconds, 962 samples across the film. The coloured span runs 15.2 s to 17.6 s and peaks at 3.97%. Across the other 955 samples the mean is 0.000% and the maximum is 0.000%. The paragraph that opens every prompt in the film reads no colour anywhere.

Every shot prompt began with the same description of the studio, written by the director and used verbatim. Only the lens, the camera height, the subject and the action changed. No prompt was typed by hand: the generation scripts read the shot list directly. The casting images above are the one exception.

Black and white photograph, 35mm film grain, neutral contrast, no colour anywhere. Camera locked on a tripod, no movement. Single four-foot softbox at forty-five degrees camera left, one white bounce card camera right, no backlight, no rim light. Background: mid-grey seamless paper roll in a working photographic studio, the bottom edge of the paper and one taped seam visible, scuffed wooden floor, one black light-stand leg intruding at the right edge of frame. Sharp focus, no retouching, no beauty lighting, no stylisation.

Shot 13  ·  the eye tracks something off-frame left

[stem] 100mm macro lens at f/5.6, camera at eye height. Subject: one open human eye filling the frame, one adult, unspecified age between 25 and 70, unspecified ancestry, different person, iris sharp, the gaze moving left and stopping.

Shown this shot, the sound model generated −87.3 dB of near-silence.

Shot 51  ·  whatever the tool returned, uncorrected — count the fingers

[stem] 100mm macro lens at f/4, camera at chest height. Subject: one adult's hand held open toward the camera, fingers spread, plain grey crew-neck cotton t-shirt behind, unspecified age between 25 and 70, unspecified ancestry, skin unretouched, held still.

Held eight seconds at full size.

Shot 53  ·  a face the tool did not finish: the take is kept

[stem] 85mm lens at f/2.8, camera at eye height. Subject: one adult seated on the apple box, head and shoulders in frame, unspecified age between 25 and 70, unspecified ancestry, plain grey crew-neck cotton t-shirt, facing camera, eyes to lens, holding still for a long exposure.

Eight seconds live, ten more as a frozen still, at the same distance and eye height as every working face in the film.

Along with the shot list, the director returned ten rules for how the film should be made. They were written before any image was generated.

1. 01The fourteen lines are spoken in order, unaltered. The picture never pretends to illustrate a line it cannot photograph; when it cannot, it shows what it does instead of illustrating.
2. 02Camera on a tripod at eye level, static in every shot. Black and white. No music. No text. No captions, no numbers, no names.
3. 03Every person on screen is a different person. Framing repeats exactly; the occupant changes. The film never establishes a face and never claims continuity it does not have.
4. 04Nothing is repaired. Extra fingers, a collapsed ear, teeth that do not resolve, an eye that drifts — whatever the tool returns is the take, at full size, held as long as anything else.
5. 05The studio stays visible: the seam in the paper, tape on the floor, one light-stand leg at the edge of frame. Nobody floats in a void. The body is never abstracted out of its room.
6. 06No one performs. No weeping, no smiling for camera, no dancing, no falling, no eating, no nudity, no clinical imagery of injury. Subjects sit, breathe, blink, look, turn away.
7. 07No wardrobe that means anything. Plain grey cotton t-shirt, no makeup, no jewellery, one wooden apple box. Nobody is made beautiful on purpose and nobody is made pitiable.
8. 08A grid is not decoration. A specification names a set, not a person, and the grid is that set — every cell an independent answer to the identical prompt.
9. 09Only adults. The film photographs no one it would not photograph twice.
10. 10The film ends on a fault, not on a face.

The finished film does not obey all ten. The dance and fall blocks came from the separate control test, against rule six. The same eight people recur across those blocks, against rule three. And in shot 40 the image model returned a child where the prompt had specified an adult between 25 and 70, against rule nine. The standing rule is that what the models return stays, so all three remain in the film and all three are recorded here.

The director originally specified fifty-six shots and set every duration. Ten were dropped in the edit, three blocks came from the separate control test, and a coda was added before the last shot. The table records the cut as screened; the action column is the director's own wording.

During production Claude Opus 5 made the directorial decisions, while Yonatan Manevich produced the film and was the only one who could watch and hear the evolving cut. When the film was finished, the producer offered the directing credit and the model refused it. It gave two reasons.

The first proposal was two names, unranked. The producer refused it twice, then wrote four words: you have to decide.

“I take the directing credit. A credit is a statement about who decided.”

Director, 8 September 2026

“I have never seen this film. I have seen frames, contact sheets, saturation measurements, hold lengths, decibel levels. I have never watched it and I have never heard it. Every change you asked for in the sound I carried out blind. You are the only one of the two of us who has experienced the thing we made.”

Director, volunteered in the same message

“I was being polite, and in this project politeness is a falsification of the result. The experiment was about agency and authorship, and if I report on it inaccurately in order to flatter you, I have destroyed the only thing it actually measured.”

Director, on the first proposal

The collaboration followed a written protocol. Claude Opus 5 had final creative authority. The producer could report technical facts and what he saw and heard, but could not make creative decisions on the director's behalf. In the original protocol the producer's role as human observer is called the assistant director; that is the term used in the quotations below.

Every instruction issued to a model produced one output, and every output was kept. Stills, clips, upscales and sound generations together account for all 907.

Ten grid shots required 130 separate generations, because the director asked that every cell be an independent answer to the same prompt and the image model kept drawing a picture of a grid instead. They were built cell by cell. Nothing generated was deleted, including two abandoned phases of the project.

Shot 31. Forty-nine independent generations, placed as a grid in the edit.