What the Model Ignores and the Ruler Cannot — Paper 03 of The Geometry of Meaning
Arc ·
Two papers ago, in the universe inside the box, I asked you to keep a hand on one thread. Kinship between stars, we said, is measured as an angle — and an angle needs a vertex, a point to stand at while you measure. Ours sat at zero: the origin of the space, the address written (0, 0, 0, …). And I asked, twice, with growing unease: measured from where, exactly? Who chose zero?
The last paper sharpened the unease into a discipline. A measurement, the surveyors taught us, is never just a number — it is a number plus the name of the frame it was taken in, and a raw number carried naked across frames is not data but noise. So we are now obligated to do something uncomfortable: take every similarity score ever read off inside the box — every retrieval, every “related item,” every semantic search that has ever ranked one meaning near another — and ask the surveyor’s question of it. What frame was this taken in? Where was the instrument standing?
This paper pulls the thread to the end. I will warn you now that the end is genuinely strange: the most widely used measurement in applied semantics turns out to answer to a landmark that the machinery being measured provably cannot see. Walking there takes five short steps — a vertex, a cone, a slide, a ruler, and a contradiction — and each one can be drawn.
First, feel how much power the vertex has. Two stars hang in the dark — fixed, serene, not going anywhere. You want to know the angle between them. Simple: stand somewhere, sight one star, sight the other, measure the spread between your two lines of sight. Done.
Now walk to a different spot and do it again.
The angle changed. Not because the stars moved — they didn’t — but because you did. Stand close to the pair and they yawn apart across half your sky; stand far away and they squeeze together into a sliver. The angle between two stars is not a fact about the two stars. It is a fact about three things: the two stars and the vertex — and the vertex’s vote is enormous. Same sky, different standpoint, any answer you like.
Inside the box, the standpoint of every angular measurement is the origin — the zero address. And here is the first thing to actually worry about: nothing lives there. Zero is not a star. No word hangs at the origin; no meaning claims that address; the butterfly has no reason ever to visit it. It is simply where the number system happens to start — the corner of the graph paper. When the box’s similarity scores are read as angles, they are angles measured by a surveyor standing at an address no one chose, in the middle of nothing, for no reason anyone can defend.
Every cosine ever taken inside the box was measured from the corner of the graph paper — and the corner of the graph paper is not a place in the sky.
“Still,” you might reasonably say, “if the stars surround the origin nicely — spread evenly around it, the way my diagrams have been drawing them — then measuring from zero is as good as measuring from anywhere.” True. And now I have to show you what real skies actually look like, because I have been drawing you idealized ones for two papers, and the real ones hold this paper’s first shock.
Chart a real embedding space — take an actual model, print the actual positions — and the stars do not surround the origin. The entire vocabulary huddles inside a narrow cone, far off to one side of zero, like a flock of birds in one corner of an enormous room. The trade calls this anisotropy — the sky is lopsided — and it is not a rare defect of badly trained models. It is the default condition of the spaces the modern world runs its searches through.
Why does it happen? The honest short version: training presses common sediment into every star at once. Frequency effects, shared context, the wholesale statistics of language — pressures that have nothing to do with any word’s particular meaning — push every vector in a common direction, the way a steady wind leans every tree on a hillside the same way. The result is that the average of the whole sky — its balance point, its centroid — sits far from zero. Remember the last paper’s revelation: the shared origin worth having was the center of mass, the point every plumb line already agrees about. The box’s sky has such a balance point too. And the ruler is not pinned there. The ruler is pinned at zero, off in the empty corner of the room, while all the meaning huddles elsewhere.
Feel the consequence with the surveyor’s instinct you now have. From a vertex far outside the flock, every star sits in nearly the same direction — so every angle between stars is small, and every cosine is large. Cat and carburetor. Justice and mayonnaise. Measured from the corner of the graph paper, everything is “similar” to everything — not because meanings agree but because the surveyor is standing in the wrong county.
Now the thought experiment at the heart of this paper — the one that turns an inconvenience into a genuine contradiction. Pick up the entire universe inside the box — every star, every constellation, the web, the butterfly mid-flight — and slide it. All of it, together, by the same shift, in the same direction. Don’t rotate anything, don’t stretch anything: one clean, uniform slide, like moving a chessboard across a table without disturbing the game.
What survives? Everything that ever mattered. Every distance between stars: unchanged — both endpoints moved together. Every constellation: intact. The web: intact. The butterfly’s flight, wingbeat by wingbeat: each displacement is a difference between positions, and differences don’t feel a common shift. The entire relational universe — which is to say, per two papers of argument, the entire meaningful universe — comes through the slide untouched.
And the box itself? This is the part that is provable rather than plausible, and though the formal proof lives in the capstone, the napkin version fits in three sentences. Everything the box does — every next-word choice, every decision that reaches the outside world — is made by comparing candidates against one another. A uniform slide adds the same amount to every candidate’s score in the comparison. And a contest where every contestant gains identical points has exactly the same winner, the same runner-up, the same everything — the addition cancels, completely, every time.
Slide the universe an inch or a light-year: the box’s behavior does not change by one syllable. The machinery is origin-blind — not approximately, not usually. Provably. Perfectly.
So the origin’s location is, to the machinery, a matter of total indifference — one arbitrary bookkeeping choice among infinitely many equally valid ones. Which would be a harmless curiosity, except for what happens when you hold the ruler up during the slide.
Run the slide again, and this time keep your eye on a cosine — the angle between cat and democracy, say, measured as always from the origin. The sky slides; the vertex, pinned to the zero address, stays where it is. Which means the geometry of the measurement changes even though the geometry of the sky does not: the flock recedes from the vertex, the sight-lines to any two stars squeeze together, and the angle between them shrinks toward nothing.
Slide far enough and every cosine in the universe is forced toward 1.0 — everything “maximally similar” to everything, an entire semantics flattened into one number — while the box, blind to the whole affair, goes on behaving exactly as before. Slide the other way and the same scores can be driven negative. The reading is arbitrary. Not noisy, not approximate — arbitrary: adjustable to any value you like by an operation that provably changes nothing the machinery does.
Now set the two facts side by side and let them grind:
The machinery is origin-blind: no measurement it makes, no decision it takes, can detect where zero is. The standard ruler is origin-obsessed: every reading it gives depends on where zero is. And the world reads that ruler by the trillion, every hour, as if its numbers were facts about meaning.
This is the drifting origin of the title, and I want to be exact about what drifts. Nothing physically slides a trained model’s sky overnight. The drift is between worlds: train the model again with a different seed, and the sediment presses in a different direction — different offset, different cone, different corner of the room. Same architecture, different layer: different cone again. Two different models — the two boxes of our whole series — two utterly unrelated cones. The offset that cosine is hostage to is set by accidents the machinery itself is provably indifferent to, and it is set differently every time. That is what it means for a reference to be unpinned: not that it wobbles while you watch, but that nothing anywhere holds it to anything — so every sky carries its own private zero, and no two of them agree.
And notice the cruelest feature of the whole arrangement, the one that should raise the hair on your arms slightly: nothing protests. When a physical instrument breaks, it usually breaks visibly — readings jump, needles peg, someone notices. This instrument fails fluently. The cosines still come back as clean numbers between −1 and 1; the rankings still print; the search results still fill the page. From inside the box there is no alarm to ring, because the box cannot feel the thing that is wrong. The wrongness lives entirely in the unexamined frame of the measurement — exactly where the surveyors taught us wrongness likes to live.
Everything so far is geometry and one napkin proof. A fair skeptic asks: does the origin’s position actually matter in practice, or is this an aesthete’s complaint? The research program behind this series put that question to the only tribunal it trusts — a pre-registered harness with the bar frozen in advance — and it asked twice, in the two directions that matter.
First, constructively. In the cross-sky alignment experiments — two independently trained models, corresponding points recovered through shared anchor stars — the single largest lever in the entire method grid was re-pinning the vertex: measuring every star’s profile from its own sky’s balance point instead of from raw zero. That one correction roughly doubled cross-sky decoding at the tested scale. Not refined it. Doubled it.
Then, destructively — the ablation. Take the full working recipe and surgically remove its corrections one at a time, bar frozen, one run. Remove the origin correction and the crowd correction together, and cross-sky structure collapses to a fraction of its corrected self. The sealed verdict reads that the corrections are not merely helpful but causally load-bearing: the alignment works because the vertex was moved, and fails when it is moved back.
I find the shape of this evidence beautiful, because it closes a loop the last paper opened. The surveyors’ shared origin was the center of mass — the point every plumb line already agreed about, found by physics rather than committee. The experiments, knowing nothing of geodesy, groped through a method grid and found the same answer: pin the ruler to the balance point of the sky. The correction that works is the planet’s trick, replayed in a universe made of meaning. And the wreckage we came to inspect is real: every raw cosine ever compared across models, every similarity threshold ever carried from one embedding space to another, was a zenith angle carried naked across frames.
But re-pinning the ruler is a correction applied afterward, by someone who knows to apply it — and that should nag at you. The planet holds its own center; nobody recomputes the Earth’s middle before every measurement. What would it take for a reference to hold itself still — and what does holding-still cost? It turns out physics has an exact answer, with a price tag denominated in joules, and engineers have been paying it, resentfully, in every circuit you own, for a century. The next paper is about grounds: what they are, what they cost, and what happens — fluently, invisibly, catastrophically — when nobody pays. The walk continues.
Refutations welcome