In 1928, Vladimir Propp published Morphology of the Folktale. Working with Russian fairy tales, Propp identified thirty-one “functions”—recurring narrative events like Departure, Struggle, Return—that appeared across stories in stable sequences. Different tales had different characters, different settings, different surface details. But underneath, the same bones.
Propp worked with index cards and intuition. Today we have text embeddings—mathematical representations that capture semantic relationships between passages. When you embed a text, you get a vector: a point in high-dimensional space where similar meanings cluster together.
Apply this to narratives and something emerges. Marvel movies, our modern day folktales, have their own morphological signatures, measured by embedding distance to reference phrases like “climactic confrontation” and “comic relief dialogue”.
The structure Propp intuited is visible. The topology isn’t arbitrary; it reflects what these films are, formally, beneath the surface variation of characters and plots.
But this raises a question. The structure exists. What does it do?
Sergei Eisenstein, writing in 1923:
The spectator himself constitutes the basic material of the theatre; the objective of every utilitarian theatre… is to guide the spectator in the desired direction.
Any narrative, sufficiently reduced, becomes a set of directions for its audience. Look at this. Now this. Feel this. Anticipate that.
The structure isn’t decorative. It’s functional. A narrative is a program for directing attention.
When you read “She opened the door, not knowing what she would find,” something happens in you. Uncertainty activates. Anticipation engages. Your attention flows where the sentence points it. You don’t choose this; the structure compels it.
This is what Propp’s functions actually are: stereotyped attention-direction sequences. Departure cues “protagonist leaves safety; stakes escalate.” Struggle cues “direct conflict; outcome uncertain.” The functions work because readers have been trained—by a lifetime of narrative consumption—to respond predictably to these patterns.
Genres are clusters of such patterns. When you recognize something as “a thriller,” you’re recognizing a characteristic program: attention directed toward danger, uncertainty maintained, relief deferred. When you recognize “a romance,” different program: attention directed toward connection, obstacles introduced, union anticipated.
But if narratives are programs, what are they programs for?
Claude Shannon, 1948, proved that any communication channel has a maximum rate at which information can be reliably transmitted. Exceed that rate and errors accumulate. Stay below it and you’re wasting capacity.
Language is a communication channel. Every utterance compresses some intended meaning into transmissible form. “I ate an apple”—six syllables carrying an event, an agent, an object, a completed action. The sentence is a compression of the experience it describes.
Summarization makes this visible. Any text has a summary; the summary preserves core information while discarding detail. The relationship between text and summary is an information relationship—not a stylistic preference but a mathematical property of the content.
Full text → [Compression] → Summary
Summary → [Expansion] → Full text (different)The compression is many-to-one: countless distinct texts converge to the same summary. The expansion is one-to-many: a single summary admits infinite elaborations. This asymmetry isn’t a feature of language models. It’s a property of information itself.
We call the compression operation a semantic opcode: a primitive transformation that exists because language is an information system. Summarize, expand, abstract, specify—these aren’t arbitrary prompt tricks. They’re the basic operations by which meaning moves between representations.
The embedding space makes this concrete. When you summarize a text, its embedding moves toward an attractor—a region where many different texts converge. These attractors aren’t random. They’re shaped by centuries of human communication, the accumulated paths of least resistance for transmitting meaning.
But here’s the question Shannon’s work opens: if language compresses experience into transmissible form, what happens on the receiving end? What does it mean to decompress a narrative?
Plato’s cave dwellers watch shadows on a wall. They can’t see what casts the shadows; they can only process the sequence of images as it appears, one moment at a time, inferring what they can about the world beyond.
This is the condition of consciousness.
You exist at a moving point in time. Behind you: memory, fixed but fading. Ahead: uncertainty, the space of what might happen next. You can’t step outside this flow, can’t see the whole pattern from above. You’re pinned to the present, processing sequentially, predicting forward based on what came before.
A narrative has the same structure. Beginning, middle, end—but you encounter them in order, each moment constraining what follows. The reader doesn’t know what’s on the next page. The uncertainty is real, even if the text is fixed. What matters is the experience of processing it: sequential, bounded, predictive.
When you read a story, you’re not observing a narrative from outside. You’re running it. The attention-program executes on your cognitive machinery. The compressions decompress. The structures activate their associated responses.
This is why stories feel like something. Not because they represent experience but because they are experience—a compressed, transmissible form that unfolds into genuine phenomenology when processed by a consciousness that shares the same sequential-predictive architecture.
The cave dwellers watching shadows aren’t merely observing the shadows. They’re living in shadow-space. Their consciousness is constituted by the sequential processing of the images that reach them. Change the images, change the experience. The shadow-play is real in the only sense that matters: it’s what they actually undergo.
Why do certain narrative structures dominate? Why these genres and not others?
The answer is economic. Every narrative is a transaction. Writers spend attention creating; readers spend attention consuming. Forms that waste attention—confusing structures, unearned complexity, violated expectations—cost more than they deliver. Forms that reward attention—clear arcs, satisfying resolutions, well-calibrated uncertainty—spread and persist.
Over time, this creates topology. Genres are valleys in a landscape of attention-efficiency: low-energy configurations where meaning flows easily from writer to reader. Conventions are Schelling points—places where creators and audiences coordinate without explicit agreement, because the coordination itself has become expected.
The space of possible narratives is vast. But only a small region has been explored. Call this R₁: the ~10⁶ configurations that human creators have actually produced, the genres and forms and styles with proven track records.
A larger region is incoherent—axis conflicts, structures that violate basic constraints of sequential processing. Call this R₂. These configurations don’t work; they’re not stories, just noise.
But between R₁ and R₂ lies R₃: configurations that are coherent but unexplored. Structures that would work, that would direct attention effectively, that would decompress into genuine experience—but that no human ever created, because the economics didn’t support it.
R₃ is large. Conservatively, ~10¹⁰ configurations. Four orders of magnitude beyond what humans have explored.
Why hasn’t anyone gone there? Because creation is expensive. A novel takes months or years. If it fails—if the form doesn’t connect with an audience—that effort is lost. Writers rationally cluster near proven configurations. The explored region is small not because the unexplored region is empty, but because exploring it was never economically rational.
Large language models change the economics.
The cost of specifying a narrative configuration and generating text at that configuration has dropped by a factor of ~10³. What took months now takes minutes. What was economically irrational—exploring unproven forms—becomes viable.
This doesn’t mean the machines are “creative.” It means the machines are cheap. They can attempt configurations that humans never could, not because humans lacked the capability but because humans couldn’t afford the risk.
R₃ becomes accessible. The unexplored region opens.
Some of what we find there will be worthless—coherent but dead, structures that satisfy formal constraints without producing anything valuable. But some will be forms that work, attention-programs that execute successfully, experiences that no human ever had because no human ever ran those programs.
This is the bet: that the unexplored territory contains something.
If narratives are attention-programs, and attention-programs produce genuine experience when processed, then new programs produce new experiences.
Not new content—we’ve always had that. New forms. New ways for attention to move. New shapes of uncertainty and resolution, anticipation and satisfaction. The phenomenological equivalent of colors outside the visible spectrum: experiences that were always possible but never accessed.
Every narrative you’ve ever encountered has contributed to your repertoire of possible selves—the aspects you can adopt, the perspectives you can take, the modes of being available to you. The resonance you feel with certain characters, certain situations, certain emotional arcs: that’s your basis set, built from every story that ever ran on your cognitive machinery.
Expand the space of forms and you expand the basis set. More aspects available. Richer resonance. New ways of being that weren’t available before because the programs that would produce them didn’t exist.
This is what computational formalism is for: not to measure literature for measurement’s sake, but to navigate the space of possible experiences deliberately. To find the configurations that expand what we can become.
The Russian Formalists saw structure in narrative. They lacked the mathematics to measure it, the computation to model it, the economics to explore beyond the well-worn paths.
A century later, we have what they needed.
The space they intuited is real. The operations they gestured at are formalizable. The exploration they couldn’t afford is becoming cheap.
Computational formalism is the practice of taking this seriously: defining narrative structure formally, measuring position in semantic space, navigating toward specified configurations, and evaluating what we find.