Generator, Not Product: How the New Disappears in Feature Space

I. The Pattern as Commodity: Generator Instead of Product
Alongside selling finished dresses to private clients, the Parisian haute couture houses of the mid-twentieth century sold the right to legally copy their models — the system of «patronage» (patrons papier). Department stores and ready-to-wear factories paid for entry to the shows, and this sum was credited toward payment for either the patterns (toiles) or the finished reference samples; retailers such as Ohrbach's and Neiman Marcus bought up entire collections in this way. By the mid-1960s, up to 60% of the $20 million annual turnover of Paris couture came precisely from the sale of reproduction rights — not a peripheral source of income but in fact the primary one, a forerunner of modern licensing. Max Meyer, a buyer for the American company A. Beller & Co., traveled to Paris 110 times over the course of his career, licensing models from Chanel, Lanvin, Paquin, Poiret, Worth, and other houses. The counterexample is telling as well: Balenciaga refused on principle to sell patterns for any money at all, and it is precisely this exception that underscores how much the norm the sale was for everyone else.
The key idea here is not historical but structural: the object of the transaction was not a thing but a generator of things. The fashion house was selling not a single dress but a generative principle from which many dresses could be derived.
II. A Modern Analogue: The Market for Instruments of Production
One need only step outside fashion to see that this structural idea did not remain confined to the mid-twentieth century. The same principle reproduces itself today in entirely different industries. The market increasingly trades not in works but in the means of producing them: Photoshop brushes (in 2017 Adobe acquired KyleBrush — a library of more than a thousand brushes by illustrator Kyle Webster — making them exclusive to Creative Cloud subscribers), color-grading LUT profiles, which colorists sell separately from the films themselves, procedural assets (Substance Designer materials, Houdini digital assets), LoRA models, which are traded on platforms such as Civitai, and processing presets, which photographers sell separately from their own photographs. The structure is the same as with the couturiers: what is sold is not the result but the generative instrument.
III. A Shift in What AI Is Trained On
And it is here that a gap arises between the market and the practice of machine learning. If the instrument of production has itself become a commodity, a natural question arises: why are models trained predominantly on finished results rather than on the instruments and the process of their application? Paintings plus brushes give a considerably fuller description of style than paintings alone — because the brush encodes not what resulted but the movement by which it resulted. It is telling that data of this kind already exists in nature: digital drawing applications (Procreate, for instance) record a full frame-by-frame replay of the drawing process rather than only the final image — that is, process data, distinct from the result, is technically already being collected, simply not yet used systematically for reconstructing an author's manner.
IV. Transferring the Principle to Fashion
Painting here is merely a convenient first example; the same line of reasoning returns us to clothing, where everything began. If the reasoning works for painting, it must work for clothing by the same logic: dresses plus patterns give a considerably fuller description of a fashion house's handwriting than photographs of dresses do. A photograph fixes only the result of the interaction of fabric, body, and gravity; a pattern fixes the intent prior to that interaction.
V. Testing the Hypothesis: Does Such a System Exist
Once the hypothesis has been formulated, it is worth testing it against actual engineering — and here a characteristic shift comes to light. Such a system already exists — but it does not do what is needed. NeuralTailor (Korosteleva, Lee, 2022, ACM Transactions on Graphics / SIGGRAPH 2022) reconstructs a 2D pattern from a 3D point cloud of a scanned dress. This is not generation but reconstruction: before the network lies one specific, physically existing dress, and the task is to strip from it the noise of reality (folds, tension, gravity) and recover the flat blueprint that lay at its foundation. The network does not care whose cut this is — an archival house's or a random home-sewn garment's: the authors explicitly declare the goal to be generalization to unseen cut topologies, meaning the model is specifically built to be blind to individual manner rather than sensitive to it. Later works (SewFormer, 2023 — predicting cut from a photograph; DressCode, 2024 — generation from a text description) expand the input modalities, but none ties generation to the handwriting of a specific author — only to the geometric class of the garment. The real test — training a model on the archive of a single house and obtaining a pattern that never existed yet is indistinguishable by an expert from a genuine work of the house — has simply not yet been posed in the published literature.
VI. Inversion of the Physical Process
Why, then, do existing systems persistently gravitate toward the generalized class rather than toward individual manner? The answer lies in the nature of the task itself. The task of 3D → pattern is an inversion of physics, not of geometry. Between the flat pattern piece and the draped form on the body stands the entire chain of physical interaction: gravity, fabric tension, friction against the body, self-contact of the folds. The formulation of the task in the original article — «disentanglement» — directly acknowledges this: it is a separating, not a projective, procedure. A multitude of different patterns can drape in ways almost indistinguishable from one another, especially where the fabric self-contacts and conceals part of the geometry from the scanner — a classic ill-posed inverse problem.
VII. The Reason for the Failure: Underdetermination
A specific mechanism of substitution follows from this ill-posedness. It is precisely because of this underdetermination that the model is forced to rely not on the data — which is insufficient to resolve the task unambiguously — but on a learned statistical prior: the frequency of panel topologies encountered during training. A prior by definition gravitates toward the frequent. This means that where the true pattern is nonstandard, the underdetermination of the task forces the network to substitute it with the nearest familiar topology, which drapes similarly but is, in essence, a different thing.
VIII. Transfer to Text Analysis
And this mechanism is not tied to fabric — it surfaces wherever intent must be recovered from a result. The same mechanism operates in the reconstruction of an argument's structure. When a model analyzes the evolution of ideas in a text with a nonstandard, deliberately unconventional composition, it tends to rearrange the elements into an order familiar for analysis — thesis-antithesis, chronology, comparison — losing the essence of the new organization of thought. The signal in favor of the genuine, unfamiliar structure is almost always weaker than the mass of texts built on familiar rhetorical templates.
IX. Semantic Collapse as a Systemic Phenomenon
The coincidence between cut and text is not accidental — one and the same structure stands behind both. The new does not disappear because of an error — it disappears because it cannot hold its ground within the existing feature space. The logic is identical in both cases: underdetermination plus a strong prior equals the systematic pulling of the new toward the familiar, and the more so, the fainter and rarer the signal of the novelty itself. The material of the inversion differs — there, fabric and gravity; here, rhetoric and generic conventions — but the structure of the failure is one and the same.
X. From the Collapse of the New to Black Swans
Having recognized this pulling-in as systemic, we run up against a question that is no longer engineering but epistemological. If the new systematically collapses into the familiar, a sharper question arises: how is the appearance of the genuinely new possible at all, if the entire mechanism of evaluation is tuned to pull it toward the familiar?
XI. The Problem Is Not in the Model but in the Human
And here it is important to shift the focus from the algorithm to the very structure of knowledge. The answer leads away from models to knowledge itself. The matter is not the architecture of the network but the inability to construct features for that which has never been observed. Specialization is an operation of narrowing the space of relevant axes down to those needed for typical work within it; consequently, a specialist by definition does not possess features that lie beyond his discipline — and no quantity of additional data within that discipline can make up for this lack.
XII. An Illustration: The Doctor and the Chair
How this lack of features looks in a concrete example is easiest to show through a single story. A useful illustration of this mechanism: a doctor received patients in his office, in one and the same leather chair. He had a group of patients with a rare illness that he could not diagnose properly. When the chair was once again sent out to be reupholstered, the craftsman noticed strange, asymmetrical patches of wear on the seat. The doctor correlated the presumed dynamics of the movements of patients with this illness when experiencing discomfort from sitting — and obtained a set of external features that the clinical picture had been lacking. The key to the solution came not from within medicine.
XIII. Bisociation as the Mechanism of Discovery
That the key came from outside the discipline is not an incidental detail of the story but its essence, which has a name. Discovery is not a deeper search within a field but a collision of different feature spaces. Arthur Koestler called this bisociation — the collision of two habitually incompatible matrices of thought, which he considered the common mechanism of both discovery and wit («The Act of Creation», 1964). The upholsterer in the example above finds the feature not because he is more perceptive than the doctor: asymmetrical wear of the leather under repeated load is not an anomaly but a central professional category of his craft. An axis invisible to the clinical ontology turns out to be axis number one in the craft ontology — and the intersection occurs because two different coordinate systems were independently looking at one and the same physical object.
And so that this does not look like the property of a single dubious anecdote, real, well-documented cases of this kind stand alongside it. The Swiss engineer George de Mestral, returning from a walk with his dog, became curious about the burrs clinging to its fur — examined them under a microscope and transferred the hook-and-loop principle into textile engineering: this is how the Velcro fastener was born. 3M chemist Spencer Silver developed a "failed," excessively weak adhesive; the solution for what to do with it came not from chemistry but from the practice of a church choir — his colleague Art Fry used the adhesive so that bookmarks would not fall out of a hymnal yet would peel off easily. This is how the Post-it came about. In both cases the decisive feature came from a domain in no way connected to the original discipline.
But a caveat about the form of the story itself matters here as well — and it returns us to how readily a prior substitutes the real structure with a familiar one. The retold cases sound like a neat chain of reasoning — noticed → correlated → constructed a feature. But this is precisely how what Nassim Taleb calls the narrative fallacy works: after the fact, a chance collision is repackaged into a neat deduction, because a coherent story is psychologically necessary, while the chaos of the genuine path is not. The real mechanism is a chance encounter of two practices unconnected to each other, and the story loses precisely what makes the case a black swan: the impossibility of arriving at it through systematic search from within a single discipline.
XIV. Taleb's Critique Becomes Insufficient
Having invoked Taleb, however, one must immediately see where his own critique stops reaching the heart of the matter. Taleb's classical critique concerns verification from outside: the observer exists separately from the system and checks whether a given model — a Gaussian one, for instance — is suited to describing a process with fat tails. Here a gap between the map and the territory is presupposed, into which one can step and compare. Applied to feature space, this gap ceases to be self-evident: if the model is not an external instrument of verification but part of the very process it is trying to describe, the question of its "applicability" becomes insufficient.
XV. A Shift in the Question: From Taleb to Soros
As soon as the external point of support disappears, the question itself changes — and along with it, the author in whom the answer should be sought changes as well. The question shifts: not "does the model fit the process" but "how fully is the model, being embedded in the process, even capable of representing it." This is no longer Taleb but George Soros's theory of reflexivity («The Alchemy of Finance», 1987): in social systems, unlike physical ones, the observer's model does not passively describe the process but itself becomes an active part of it — the object shifts from the very fact of being observed and thought about, and therefore there is no fixed "true" process to which the model could ultimately draw closer. It follows that the fullness of understanding is fundamentally unattainable not because of a lack of data — that would be Taleb's problem — but because the very act of understanding is an operation unfolding inside the same reality it is trying to encompass, and therefore never concludes from outside it.
XVI. A Gap in the Sources as an Occasion for a Broader Question
This embeddedness of understanding within reality has an unexpected practical trace — and it comes to light precisely where we attempted to verify one of the stories against the sources. The story of the chair is not to be found in the sources — not because it did not happen, but because a search uncovers only what has already passed through the bottleneck of publication, cataloguing, digitization. Absence from the search results is not absence from reality but the boundary of the documented surface. From here a natural transition follows to a broader question: to what extent, in general, is reality — or a person within it — documented at all?
XVII. The Archive as a Thin Projection of Reality
The answer to this question repeats — now at the level of the archive — the same figure we saw with the model. Documents are not reality itself but a thin surface remaining after an enormous number of selection filters. This is the same structure as the model's blindness to a feature outside its space of axes, only at a different level: a search index does not see an event if it has not entered its space of sources.
XVIII. An Unexpected Thesis: A Scholar Is Easier to Erase
If the archive is a thin, selected surface, a counterintuitive conclusion follows from this about whom it is easier to remove from that surface. From this follows a counterintuitive thesis: a scholar is easier to erase from history than an ordinary person.
XIX. The Reason: Reality as a Network of Couplings
The paradox dissolves as soon as one looks at how exactly the trace of each is distributed. The public identity of a scholar is concentrated in a few centralized, thin artifacts — the named authorship of publications, institutional records, portraits, correspondence — which are controlled by a small number of gatekeepers: journals, academies, the state. A classic example is the Roman damnatio memoriae: after Caracalla killed his brother and co-ruler Geta in 211, the senate ordered his name struck from inscriptions and his face erased from the family portrait — the famous Severan Tondo, where the scraped-out section is still visible today. Soviet practice of the same order: Nikolai Yezhov — ironically, the People's Commissar for Water Transport — was cut out of a photograph on the Moscow-Volga Canal and replaced with water; Trotsky was scrubbed from hundreds of photographs beside Lenin. David King documented this industrial scale of state retouching in the book «The Commissar Vanishes» (1997). And within science, it is enough to recall the fate of Nikolai Vavilov: the leading Soviet geneticist was arrested and died in prison in 1943, while his institute and genetics as a discipline were discredited for decades and pushed out of official science under Lysenko — centralized suppression through control of only a few nodes (the academy, journals, textbooks).
An ordinary person is structured differently: he is deeply embedded in an enormous number of heterogeneous, redundant social ties — records of birth, marriage, property, taxes, court proceedings, the memory of neighbors, letters, family photographs, distributed across a multitude of households. No censor controls all of these nodes simultaneously. An example from microhistory is telling: Carlo Ginzburg, in the book «The Cheese and the Worms» (1976), reconstructed the worldview of Menocchio — a sixteenth-century miller unknown to anyone — four centuries later, relying on the records of the inquisitorial court. Even a truly ordinary person leaves a trace dense and redundant enough to be reconstructed — because he was simultaneously inscribed in parish, tax, and court records. Reality, in other words, is a network of couplings, and it is precisely the density of this network, not fame, that determines resistance to erasure.
XX. Abandoning the Map-and-Territory Metaphor
To accurately describe everything said above — from cut to archive — the familiar pair of concepts no longer suffices. It is important here to abandon one imprecise comparison: what is happening is not a story about the map and the territory in the classical sense, where there is one map and one territory, differing in scale.
XXI. A Three-Level Scheme
Instead of two levels, a third is needed — and then everything examined above falls into place. A three-level scheme describes what is happening more precisely. Reality is a topographic map: it contains the entire relief, the full plenitude of detail prior to any selection. The model is a road map: a designed abstraction that selects part of the features as significant (roads) and discards the rest — by design, for a specific task of use. Our representations are a route map: yet another reduction laid over the road map — one specific path traced for one specific trip, even more selective than the road map itself.
XXII. A Consequence: Errors at Different Levels of Reduction
And it is precisely because of these three levels that all the cases examined turn out to be manifestations not of one but of errors different in nature. It follows that errors can arise at different levels of reduction, and these are errors different in nature. Some arise already at the construction of the model — at the decision of what to count as a "road" at all, that is, at the selection of features: this is exactly the underdetermination that forces NeuralTailor to pull a rare cut toward a typical topology, and text analysis to rearrange arguments into a familiar order. Others arise already at the use of the finished model — at the choice of which route to trace over the road map, that is, at interpretation and application: this is a separate, additional act of selection, layered atop the first. Eliminating the error at one level — more data, a better prior, a more expressive architecture — does not touch the error that arose at another. This is precisely why one and the same result — the disappearance of the new, blindness to a feature outside the observer's ontology, the thinness of the archive relative to reality — can be reached on entirely different floors of reduction, and no local repair of one floor guarantees the integrity of the entire structure.