A Study in Character Interiority, Continuity, and Conscious Assembly
What does it take for a generated person to become a character?
Not simply a face that remains recognizable across shots. Not a reusable avatar. A character: someone who seems to carry a continuity of self from one moment to the next, whose gestures suggest an inner state, and whose presence implies a life extending beyond the few seconds in which we see him.
That became the question behind Study 32 of the Life-world Series.
The resulting film, Errand, is only about twenty seconds long. A young man leaves a neighborhood grocery store carrying a paper bag. He walks through a damp city street at dusk. Eventually, he reaches a rooftop where a small film crew is preparing to shoot. What first appears to be an ordinary grocery run turns out to be something else: he has been bringing supplies to the production.
That is almost the entire plot.
But the experiment was never really about the errand.
It was about Eli.
From the object to the person
Study 31, Four Witnesses, held a single antique pocket watch constant while changing the world around it. Relic. Specimen. Breach. Vessel. The object remained, but its felt meaning changed as the environmental and authorial layers around it changed. In the second edition of AI Cinematic Realism, I described the study as a demonstration of conscious assembly: the perceptual surface stays comparatively stable while the world and its meaning are deliberately transformed.
Study 32 shifts the experiment from the object to the person.
This time, the question was not whether an object could retain its identity while its meaning changed.
It was whether a synthetic person could retain enough continuity to begin acquiring one.
That distinction matters within AI Cinematic Realism. The Ideational Frame identifies character interiority as the sense that behind a figure’s face is intention, feeling, a life. Together with narrative implication, it belongs to the authorial stratum: the layer in which story and subjectivity are shaped.
A consistent face is not yet a character.
The harder achievement is a continuity of self.
Building Eli before building the film
Errand was created entirely in Google Flow, Google’s AI filmmaking environment. The production used Flow not simply as a text-to-video generator but as a system for constructing a reusable character, generating individual shots, iterating on failed outputs, and assembling the resulting material into a short sequence.
That production context matters.
Rather than beginning with an image, environment, or visual transformation, I first created a reusable character in Flow.
I named him Eli.
His visual reference established a face, clothing, and physical presence. A fuller body reference helped establish his silhouette and wardrobe. A voice was also assigned, even though the final film ultimately did not depend on dialogue.
But appearance was only the beginning.
Drawing from the Character Logic section of the AICR Production Manual, I also established how Eli should behave:
Quietly observant and self-possessed. His expressivity is restrained and natural rather than theatrical. Reactions often begin in the eyes before reaching the mouth; smiles are small and sometimes asymmetrical. His gestures are economical and purposeful, with relaxed shoulders and a grounded posture. He moves at an unhurried, deliberate pace and tends to pause and observe before acting. When uncertain, he becomes still rather than fidgety. His behavior should remain subtle, believable, and consistent across scenes.
The Production Manual separates character construction into Character Arcs, Emotional Grammar, and Performance Constraints. Emotional Grammar includes emotional range, micro-expression tendencies, gesture vocabulary, posture norms, and movement style. Performance Constraints specify expressivity level, realism versus stylization, physical limitations, and behavioral consistency.
The perceptual craft section carries those decisions further into Synthetic Performance: gesture vocabulary, micro-expression grammar, posture and embodiment, and movement style.
That distinction became unexpectedly important.
The question was no longer simply:
Can Flow reproduce Eli?
It became:
Can Eli continue behaving like Eli?
That is a much harder problem.
The stubbornness of matter
The production quickly exposed why synthetic filmmaking cannot be reduced to writing increasingly elaborate prompts.
In an early version of the first shot, Eli was supposed to leave the grocery store carrying the paper bag, answer a phone, and continue walking.
Nothing about that action is remarkable.
In photographed cinema, the physical world solves most of it automatically. A person has two hands. A bag has weight. A phone occupies one hand. Gravity remains stubbornly committed to its job.
The generative model had other ideas.
Across repeated attempts in Flow, Eli dropped the bag when he answered the phone.
I changed the choreography. I placed the phone in his hand from the beginning. I specified how the bag should be carried. I explicitly instructed that he maintain his grip.
The bag kept falling.
Eventually, I removed the phone.
That was not surrender.
It was direction.
AICR places material plausibility in the Ideational Frame: surfaces and objects must behave according to their implied nature. At the perceptual stratum, motion must carry weight and momentum; the body, its clothing, and the objects it touches must seem to belong to the same physical event.
The paper bag became an unexpectedly useful reminder that realism is not a matter of resolution. A model can render skin, fabric, wet pavement, storefront light, atmospheric depth, and an apparently convincing city street while struggling with the logic of two hands and one bag.
Local plausibility is not the same thing as coherence.
The vegetables that stayed
A quieter problem survived into the finished film.
When Eli first leaves the grocery store, a few vegetables protrude from the open paper bag. They establish an expectation about what the bag contains.
Later, on the rooftop, the contents are revealed as supplies for the film shoot.
The vegetables are gone.
This was not an intentional transformation.
I tried repeatedly in Flow to regenerate the opening so that the bag remained visibly open while its contents stayed below the camera’s view. The distinction seems simple: an open paper bag without anything protruding from the top.
The model resisted it.
Leave the bag open and groceries appeared. Alter the prompt too aggressively and other parts of the shot changed. Even explicit instructions about the camera angle and visibility of the contents did not reliably produce the desired result.
Eventually, I stopped trying to force a correction that the generation process would not reliably give me.
The discontinuity remained.
That decision is important because conscious assembly does not mean perfect continuity at any cost.
The AICR framework describes conscious assembly as the condition in which coherence that a physical world once supplied automatically has to be assembled — or accepted — on purpose.
But that does not mean every retained mistake becomes art.
The vegetables are not an authored glitch.
They are a continuity fracture.
Noticing an artifact does not retroactively make the artifact intentional.
What becomes intentional is the decision about what happens next: regenerate it, restructure the sequence, hide it, foreground it, or determine that the film can sustain it and let it remain.
That is different.
The dropped bag broke the physical action badly enough that I redesigned the shot.
The disappearing vegetables create a smaller continuity problem across shots, one I recognized and knowingly allowed into the finished film.
This distinction matters because otherwise “conscious assembly” could become an excuse for carelessness. The AICR framework draws the line: glitches that break structural commitments—such as a face losing continuity or established physical logic contradicting itself—remain failures rather than automatically becoming expressive texture.
Intentional imperfection requires actual intention.
Sometimes accountable authorship means correcting the machine.
Sometimes it means admitting where you could not.
Generate. Evaluate. Correct. Regenerate.
Errand was not produced by discovering the perfect prompt.
It emerged through the generative loop.
The AICR Production Manual formalizes that loop as:
Generate → Evaluate → Correct → Regenerate
Generation produces material.
Evaluation examines that material across the perceptual, environmental, authorial, and sonic layers.
Correction identifies the failure rather than merely calling the output “bad.”
Regeneration tests the correction.
The loop exists because intention does not transfer perfectly into generation.
Some constraints are relatively text-legible. Others are reference-legible: character identity, facial structure, voice identity, or a specific visual object cannot always be reliably preserved through description alone.
Still others—theme, narrative arc, emotional rhythm—cannot simply be transmitted to the model and must instead be enforced through evaluation and postproduction.
The Manual describes the loop as the discipline that recovers what this translation loses.
Errand made that principle tangible.
I generated.
Eli became someone else.
I corrected.
His identity became more stable, but performance drifted.
I corrected again.
The bag misbehaved.
I simplified.
One shot maintained the character but created a continuity problem in an object.
I evaluated whether that problem destroyed the film.
It did not.
The finished twenty seconds were never waiting inside a prompt to be discovered.
They were negotiated into existence.
Three strata, one very small errand
Because the finished film is modest, its AICR structure is easy to overlook.
That modesty is precisely what makes it useful.
Perceptual: does Eli remain embodied?
At the perceptual stratum, the film depends on ordinary things behaving ordinarily.
Eli needs weight.
The bag needs weight.
Walking needs momentum.
Clothing needs to respond to the body wearing it.
Wet pavement needs to belong to the light falling across it.
None of this is spectacular.
That is the point.
The perceptual stratum operates at the speed of the eye, where the viewer accepts or rejects motion, material, and embodied presence before analysis begins. A highly polished render can still fail when surfaces forget their weight or movement loses physical conviction.
The bag problem mattered precisely because the rest of the shot could look convincing.
Environmental: does the journey belong to one world?
The street sequences establish a damp early-evening city: muted daylight, wet pavement, storefront illumination, pedestrian movement, a quiet rhythm of ordinary life.
The rooftop changes the location without abandoning the emotional world.
AICR calls this atmospheric continuity: the persistence of mood, tone, light, and ambient feeling across separate images. It is one of the Four Pillars of conscious assembly because separate synthetic generations do not inherit a physically continuous world behind them.
The connective tissue has to be constructed.
The rooftop does not need to look like the sidewalk.
It needs to feel as though Eli could have arrived there from it.
Authorial: why is he carrying the bag?
The final shot changes the meaning of the earlier ones without negating them.
A man carrying a paper bag leaves a grocery store.
We know that visual grammar.
We fill in the obvious story ourselves: groceries, home, routine.
Then the rooftop reveals the film shoot.
The bag was part of an errand, but not quite the errand we assumed.
That small recontextualization gives the preceding shots narrative implication. Eli was going somewhere. The bag mattered. His movement had an objective even when the film withheld it.
The story is tiny, but there is an important difference between a tiny story and no story at all.
The authorial stratum is where AICR asks whether a work feels meaningfully authored rather than merely generated. Character interiority is especially demanding because no amount of visual fidelity can create it by itself.
The synthetic face has to read as the outward edge of an inner state.
Eli’s interiority in Errand is deliberately modest.
He does not cry.
He does not deliver a monologue.
He does not undergo a dramatic transformation.
He simply seems to have somewhere to be.
Sometimes that is enough to begin a life.
Conscious assembly without a master plan
There is an important qualification.
Errand was not produced by following the AICR Production Manual from its first preproduction step through its final audit.
The full Manual begins earlier. It asks the filmmaker to articulate an Authorial Stratum before generation: thematic spine, narrative commitment, genre grammar, visual style, world rules, character logic, and supporting authorial documents.
Errand was more exploratory.
It began while I was developing synthetic-video material for a presentation. Eli came before the completed story. The situation evolved through production. Possibilities emerged, failed, were rejected, simplified, or redirected according to what the generated material could actually sustain.
Google Flow was significant here not because the study seeks to validate a particular platform, but because its affordances shaped the form that exploration took.
The ability to construct Eli as a reusable character made continuity more achievable.
The repeated generations exposed where that continuity remained fragile.
The platform made some forms of control surprisingly accessible while resisting instructions that seemed almost trivial.
Those constraints became part of the filmmaking process.
Another model, another interface, or another generation of Flow might behave differently.
The production lesson remains.
Study 32 therefore became a test of whether AICR can function not only as a preproduction architecture but also as a diagnostic discipline during discovery.
Here, it did.
Character Logic gave Eli behavioral boundaries before the film knew exactly what it wanted him to do.
The three strata made it possible to distinguish an identity problem from a material problem, an environmental problem from an authorial one.
The generative loop turned rejection into method.
And conscious assembly became the process of deciding which generated possibilities belonged to the film and which did not.
AICR describes the AI filmmaker’s condition plainly: a camera once inherited a coherent physical world. The synthetic filmmaker inherits no such guarantee. Whatever structural and emotional logic the work needs must be assembled—or accepted—deliberately.
Errand was assembled in exactly that sense.
Not everything was planned.
Everything that remained was chosen.
The character is not the face
The most useful lesson of Study 32 may be the simplest:
Character consistency is not identity consistency.
A generator can reproduce the same face and still lose the character.
Eli became convincing only when several kinds of continuity began reinforcing one another: facial identity, clothing, posture, movement rhythm, restrained expression, environment, pacing, and the implication that he was acting toward an objective.
The Production Manual treats character continuity as more than facial resemblance. Its perceptual workflow asks for consistency in facial structure, hair behavior, clothing physics, and gesture grammar, while Character Logic establishes emotional grammar and behavioral consistency before those images are generated.
Study 32 suggests a useful way to state the difference:
A reference can preserve what a character looks like. Authorship must preserve who the character seems to be.
That “seems” matters.
Eli has no consciousness behind the image.
There was no actor living this errand.
No remembered rooftop.
No experience of wet pavement.
No private anticipation of reaching the shoot.
And yet the film asks us to infer intention.
That is precisely the tension at the center of AI Cinematic Realism.
Synthetic cinema contains no recorded event behind the image, but it can still generate narrative implication, affect, and the sensation of presence.
Character interiority, in the framework, is not proof that a synthetic person possesses a mind. It is the degree to which the work successfully constructs the sense of one—a continuity of self that no render can supply by itself.
The goal is not to pretend there was a life behind the lens.
There was no lens.
The question is whether authorship can give the synthetic figure enough coherence that, for twenty seconds, we perceive a life anyway.
From Four Witnesses to Errand
Study 31 asked what happened when one object passed through four meanings.
Study 32 asks what happens when one synthetic person passes through several moments and is expected to remain someone.
The shift is small, but it moves the Life-world Series more directly into the authorial stratum.
The challenge is no longer simply to make Eli recognizable.
It is to make recognition matter.
A paper bag, a wet sidewalk, a subdued expression, a rooftop, a camera waiting to be used—none of these contains meaning on its own.
The model can generate all of them.
The film exists in their relationship.
And perhaps the disappearing vegetables belong there too—not as a triumph of imperfection, but as evidence of the negotiation itself.
The world does not always hold.
The model does not always obey.
The filmmaker cannot always repair everything without breaking something else.
AICR does not require hiding that condition.
It requires answering for it.
That is where authorship remains.
Images can be generated.
Cinema must be authored.
And perhaps a synthetic character begins to live—not when the machine finally remembers his face, but when the filmmaker gives us a reason to wonder where he is going.
Joni Gutierrez, Ph.D. 
Leave a comment