The Framework Told Us Our Test Was Wrong
I bought Bertrand Russell’s Principia as a young man and never read it. Fifty years later it was Type Theory — Russell’s own repair for his own paradox — that finally pulled me in, and within a week it handed me a correction to a test we had written wrong: in a world with no distance, a polygon is a circle.
When I was young I collected heroes, and I am no longer certain why I chose the ones I did. Hermann Buhl, who walked off Nanga Parbat alone. Einstein and Lincoln and Mark Twain and Pelé — the usual crowded shelf of a boy who read whatever he could find by them or about them. One of them was Bertrand Russell, and at some point I acquired his Principia Mathematica and set it on a shelf, the way you set a mountain on a list, certain that one day I would climb it. I never did. It sat there for the better part of a lifetime being admired and not read.
Category theory was not taught to undergraduates when I was one, in the early 1970s, and type theory as we now know it did not yet exist. So this past year, learning both from scratch, I have been slow — pausing for weeks at a time to let an idea settle. What finally walked me through Russell’s front door was not the Principia on the shelf but Type Theory itself, which grew directly out of the theory of types Russell invented to escape his own paradox. Fifty years late, I met my hero by learning the tool his failure forced him to build.
And within about a week, that tool told me our own test was wrong.
A polygon is a circle
We had been checking whether the lattice treats every direction the same — whether it is isotropic. The check we wrote measured the lattice’s propagation and asked whether it came out round: equal in all directions, the way a circle is equal in all directions. It did not come out round, so the test reported the lattice as lopsided, anisotropic, biased toward certain directions.
I objected, and the objection came straight from what I had just been reading. The “circle” you are comparing against is a Euclidean circle — round because Euclidean distance is round. But the lattice has no Euclidean distance. It has only adjacency: this point touches those points, and nothing in the substrate says how far apart anything is. If you insist on a notion of distance anyway, the most honest one is the taxicab distance — the number of steps along the grid, the way a taxi crosses a city that has no diagonals. And here is the thing that stopped me: the unit circle of taxicab distance is not round. It is a diamond. A polygon. The set of all points “one unit away” is a square standing on its corner. In a world whose only distance is the taxicab, a polygon is a circle.
So our test had been holding up a round ruler and marking the lattice down for failing to match it — but the round ruler was never ours to hold. It was imported from a continuum the lattice does not contain. Calling the lattice anisotropic was smuggling in a measurement no observer living inside it could ever make.
The criterion was wrong, and we rewrote it. The corrected version says something sharper and, I think, more honest: absolute isotropy is not a testable thing at all. If the lattice distorts every direction, but light and matter and rulers and clocks all distort together in exactly the same way, then no experiment can see the distortion — there is nothing left to compare it against. A bias shared by everything is invisible. What is physical is not whether one sector looks round, but whether two sectors disagree. Establishing that light moves in a certain cone tells you about light; it does not tell you that matter shares the cone. Only the disagreement between them could ever be measured, and only the measurable is physics.
When we re-scored our own tests under the corrected standard, results I had been ready to count as passes quietly became partials — not because anything broke, but because a test that establishes one sector had been mistaken for one that establishes agreement between sectors. That is exactly the kind of demotion you want a method to be able to force on you.
This is not new mathematics. Poincaré said it more than a century ago: geometry is partly a convention, and a distortion you cannot detect is not a fact about the world but a choice of description. What was new, for me, was arriving at it from the inside — being handed the correction by the very formalism I had picked up to understand my old hero.
An honest ledger
It would be easy, and dishonest, to tell this as a story about category theory solving our problems. It did not. But it paid three real dividends, and they all fell on the same axis, so the pattern is worth stating plainly.
It gave us the line between what is real and what is bookkeeping. In category theory there is a precise notion — naturality — for a transformation that acts the same way on everything at once. The theory’s own theorems say that such a uniform transformation carries no observable content: nothing you can honestly measure distinguishes a thing from its uniformly-transformed self. A distortion that touches every sector identically is exactly this kind of transformation, and that is why it is invisible. The observable content of a theory is precisely what fails to transform uniformly — the part that refuses to be waved away. That single idea is what forced the isotropy correction above. It is the physicist’s distinction between a gauge choice and a real observable, stated in the language of the mathematics.
It made the inside-observer respectable. There is a foundational result (the Yoneda lemma) whose moral is that a thing is completely determined by how it relates to everything else — you never need to open it up and inspect its “essence.” For a framework whose whole predicament is that its observers are trapped inside the system, probing it only through relations, this is liberating rather than paradoxical: measuring the universe from within recovers its structure faithfully, up to the one thing you can never get — the essence behind the relations. That is not a limitation we imposed; it is one the mathematics says is intrinsic.
And it caught us in an outright error before it became a claim. At one point we were tempted to model the lattice as a single global roll-out over all its possible configurations — enumerate the arrangements, sum over them. The sheaf-theoretic account of quantum contextuality showed that this framing is, in disguise, a hidden-variable model — the kind that provably cannot violate the Bell inequality. Had we built on it, we would have quietly placed ourselves inside the Bell bound, contradicting one of the results the framework most needs to keep. The formalism caught the mistake while it was still a temptation.
Now the other half of the ledger, which matters more because it is the half one is tempted to hide. Category theory did not do the geometry. Counting the faces of the structures we work with is ordinary combinatorics — binomial coefficients. The propagation results are linear algebra. The arguments about which rotations survive are plain group theory. The constraints that pick out particles are physics. Category theory framed the questions and corrected our criteria; it did not generate a single one of those answers. Anyone who tells you the abstract machinery derived the content is selling something.
So the ledger closes on one line. Category theory paid on the question of what counts as observable — separating real structure from artifacts of how we chose to describe it — and it did not, and does not, manufacture content. That is a narrower claim than the enthusiasts make and a larger one than the skeptics allow, and it happens to be the true one.
A prediction with no dial
There is a payoff from all of this that stands entirely on its own, and it is the one I would hand a skeptic first.
Calibration — the subject of an earlier essay — supplies only the dictionary that turns the lattice’s counts into worldly units: what a setting is in radians, what a tick is in seconds, what a spacing is in metres. But the sharpest tests in quantum mechanics are correlations, and a correlation is a pure number — dimensionless, unit-free. It follows immediately that no calibration can move a dimensionless prediction. There is no constant anywhere in the dictionary that touches it; there is nothing to turn.
That is a strong place to stand. If the framework computes a particular number for a Bell-type correlation and the laboratory reports a different one, there is no knob I can quietly adjust to reconcile them — the prediction is immune, by construction, to the oldest objection in theoretical physics: “you just fitted it.” A theory that cannot be tuned to agree can only agree or fail, and that is the only kind of agreement worth anything.
The cave, with the chains rewritten
For those who have followed the program’s Plato’s-cave framing, the same idea recolours the old allegory. The prisoner is not chained to a wall by force. He is held in place by invariance — and being himself one of the invariant things, he has no direction in which to turn, because turning would be a distortion shared by the turner and everything he might turn toward. Two things change from the original story, and both are gentler. The shadows are not degradations of a brighter truth; they are a selection — a true subset of what is there, honestly seen, not a deception. And a filter, unlike a wall, has a mesh — a structure that can in principle be measured from inside, by an observer patient enough to map what does and does not get through.
What the method is for
I set out this year to finally understand a hero, and what I got instead was a correction to my own work, delivered by the mathematics he was forced to invent. There is a symmetry in that I did not expect. The last time the program published a correction, it was a physical test telling us that a construction we had built did not survive contact with the sky. This time it was a piece of century-old mathematics telling us that a test we had written was asking an unanswerable question. In both cases the same virtue was at work: a way of doing the work that is able to turn around and tell its own authors they were wrong — and does.
That is a claim about how we work, not about what the universe is. Which geometry the framework will settle on, what the successor architecture looks like, whether the pieces fit — all of that is unfinished, and I will not pretend otherwise here. The only thing I am reporting is a discipline: hold up no ruler the lattice does not own, count as real only what two sectors could disagree about, and let the formalism correct you when it can. Fifty years to open a book, and the first thing it did was hand me back a red pen.
Related: “Counting Is a Discovery. Seconds Are a Decision.” on why the counting engine holds no imported constants; “The optical axis, tested” on the single-domain no-go; and the program’s Statement of Intent on the expressive-power-not-ontology lane this essay stays inside. Claim-by-claim status lives in the claim map.