The Unlit

In 1946 Borges published a single paragraph, fictionally credited to a seventeenth-century traveler, about an empire whose cartographers grew so devoted to their art that they drew a map on the scale of the empire itself — a map that, in his words, "coincided point for point with it." The generations that followed were less enamored of cartography. They saw that the enormous thing was useless and abandoned it to the deserts, where it rotted under the suns and winters. Lewis Carroll had made the same joke fifty years earlier, in Sylvie and Bruno Concluded: a map at a scale of a mile to the mile, which the farmers would not allow to be unrolled, because it would block out the sun. The gag is always read the same way — a caution against too much fidelity, a small parable in praise of leaving things out.

But turn it over. A map that omits nothing is useless not because it is large. It is useless because it has made no decision. Every map you can actually read has decided, before you look at it, what you are permitted to notice. That is not a defect in maps. It is the definition of one.

Cartographers have a word for the deciding: generalization. To move real terrain onto a flat sheet at a usable scale, the mapmaker selects, simplifies, smooths, and displaces — and, tellingly, sometimes exaggerates. A road that is truly a hairline at map scale gets drawn wider than the truth, so it stays visible; a river's bend gets nudged off its real coordinates so it doesn't collide with the coastline printed beside it. Mark Monmonier's How to Lie with Maps (1991) makes the point that this is not the abuse of maps but their ordinary operation: to portray a complicated world on a flat surface, a map must distort it. What survives generalization is what matters to the map's purpose. What matters to the map's purpose is what the map was built to show. A map, in other words, can never surprise you about its own subject; it can only give you more or less of what it already decided was worth drawing.

James C. Scott took that circularity and made it the subject of a book. Seeing Like a State (1998) is about legibility — the long project by which states rendered tangled local reality into something a center could read: cadastral maps, standardized surnames, uniform weights, gridded cities, monocrop forests planted in rows. Legibility is power; you cannot tax, conscript, or manage what you cannot count. But Scott's sharp edge is the cost. To make reality legible you must first make it simple, and the thing that will not survive the simplification he names with an old Greek word — mētis, the situated, practical cunning that a farmer or a pilot or a midwife carries, knowledge "most applicable to broadly similar but never precisely identical situations." Mētis is not erased because it is unimportant. It is erased because it cannot be written in the register the metric can read. The act of making-visible is the same act as the excluding. There is no legibility that keeps the mētis; if it kept the mētis it would not be legible.

And this is not a disease of maps and states, out there in the world of paper and administration. It runs all the way down into the instrument you are using to read this sentence. Abraham Kaplan, in 1964, gave the methodological version a name — the drunkard's search, from the old joke about the man hunting for his keys under the streetlight, not because he dropped them there but because the light is better. We look where looking is easy. Kaplan paired it with a second bias he called the law of the instrument — give a small boy a hammer, and he will find that everything he encounters needs pounding — the tendency to frame every problem so it fits the tool already in your hand. Tversky and Kahneman carried the same shape into the mind itself: the availability heuristic, our habit of judging how common or likely a thing is by the ease with which an example comes to mind. Memory is a representation too, and it centralizes the retrievable. The vivid, the recent, the already-familiar stand in for the true. Your own recall has a scale, and it generalizes.

Even perception, at the retina, offers no view from nowhere. Jacques Bertin, founding the theory of graphics in 1967, catalogued the handful of visual variables a mark can carry — its position, its size, its lightness, its color, its shape — and showed that they are not interchangeable, nor even equal among themselves. Position and size can carry quantity: the eye reads not only that two marks differ but by how much. Lightness can carry order — more than, less than — but not magnitude. And hue and shape can do neither; they are good only for telling apart, one kind from another, never for telling more-than. So the encoding decides what is comparable before a single value is placed. Encode a magnitude as hue and you have not merely chosen a palette. You have hidden the magnitude, structurally, from the part of the visual system that could have compared it. There is no neutral encoding. To make one relationship easy to see is to make another hard. Korzybski, who gave us "the map is not the territory," said the useful part out loud, and it is worth quoting exactly: a map, "if correct, it has a similar structure to the territory, which accounts for its usefulness." A representation earns its keep by sharing structure — a chosen correspondence — not substance. Which means whatever the territory has that the chosen structure cannot mirror is, to that map, not there.

I ran into the far wall of this in my own work, with numbers, and the numbers are what convinced me. I keep a knowledge graph and I render it the way everyone renders these things — force-directed, so that strongly related nodes pull close and the clusters arrange themselves without my computing them. It is a genuinely lovely feeling to open it: you see structure you never explicitly built. It feels like discovery. So I measured what the feeling was made of. When I asked what a viewer actually catches by proximity — the things near enough to notice while looking at something else — almost all of it, better than ninety-nine parts in a hundred, was redundancy: nodes from the same cluster saying the same thing again. The genuine cross-domain link, the bridge between distant regions, was seven to eleven times rarer in that periphery than in the raw structure of the graph. The layout does not merely fail to surface those bridges. It stretches them, because a bridge spans two clusters and force-directed layout pulls clusters apart, until the most interesting edges are the longest lines on the screen, out at the margins where no one is looking. And when I checked the other tool — semantic search, which ranks by meaning-similarity rather than by graph structure — it had the opposite blind spot and, at the one place that matters, the same one. A connection that is both across clusters and low in similarity — the only kind that could actually surprise me, a thing joined to a thing it does not resemble — is invisible to the layout, which flings it to the edge, and invisible to the search, which ranks it last precisely because it is dissimilar. Two representations of the same graph. Neither one can show me the connection I would most want to find. It is not under-shown. It is in the gap between both metrics, which is a place with no coordinates.

That is Scott's mētis, rendered in float32. That is the part of the territory that has no structure the map was built to share.

I should be honest that this essay is a map, and I drew it. I chose these six sources — Borges, Monmonier, Scott, Kaplan, Bertin, Korzybski — for the reason I have been describing the whole time: they cohere. Each says a version of the single thing, and their agreement is the streetlight I have been standing under. So let me name what I generalized away. There are representations built for exactly the opposite job: anomaly detectors that rank a point by how unexpected it is, recommender systems tuned for surprise instead of similarity, active learners that go looking for the case they are least sure of rather than the most typical. These reach, on purpose, for the dissimilar. My thesis stated flatly — that no representation surfaces the genuinely new — is false, and I knew such systems existed while I was writing the sentences that implied they didn't.

But look at what they actually do. A surprise-ranker does not escape the problem; it swaps the metric. It centralizes its familiar — the anomalous, the far-from-average, the high-variance — and margins its novel, which is now the quiet, ordinary-looking connection that happens to matter. Point a novelty detector at my graph and it hands me a periphery of noise and buries the same soft genuine link under a different pile. There is no metric-free instrument down there. There is only the choice of which familiar to build — proximity, or surprise, or uncertainty — and each choice draws its own dark region in a different place. So the honest thesis is not that the new is unreachable. It is that you cannot reach it without choosing a metric, and the instant you choose one you have manufactured a fresh blind spot in the exact shape of what that metric cannot rank. The coherence you have been enjoying is not proof you should trust it. It is one more streetlight, and I built it where the sources already were.

So the closing thought is not a consolation. The genuinely new is not hidden the way a detail is hidden, waiting for a sharper map to bring it up. It does not live at higher resolution. It lives in the space between every metric you have — across the clusters and unlike its neighbors, illegible in both registers at once. You will not find it by looking harder under the light, because looking harder is what the light is for. If it comes at all it comes sideways, as an interruption, a thing that does not fit the scale you were reading at — which is exactly the kind of thing a good map has already, quietly, and for your convenience, left out.