PROJECTIVE CORRELATION THEORY

Start with the one move everything else follows from. Ordinarily, you are handed a space, you put things in it, and you then ask how strongly those things are related. PCT runs that backwards. You are handed only the relatedness — a single number for every pair, saying how much knowing about one tells you about the other — and distance is defined out of it. Two things that are strongly correlated are, by definition, close together. Two things barely correlated are far apart. Geometry becomes a summary of correlation rather than a stage for it.

The word pregeometric in the paper means exactly this: the starting objects have no geometry, and any geometry that appears has to be earned. The word projective refers to the smoothing step — a projection that turns the raw correlation structure into something an observer could actually measure. And kinematic, which the paper repeats insistently, means the framework describes what a correlation structure implies and not how one evolves. There is no equation of motion here. That is a limitation the paper states about itself, not a criticism from outside.

So the shape of the thing is: four primitive objects, five conditions saying when geometric language is allowed, one theorem converting correlation into a metric, one derived quantity that measures how many dimensions the result has at each scale, and six registered predictions with numbers attached. The next four sections take those in order. What it is not is a theory of everything — it has no matter, no dynamics, and it says so on its first page.

One consequence is worth flagging early because it recurs throughout. If distance is derived from correlation, then strong correlation between two distant regions is not a paradox to be explained — it is a statement that they were not far apart in the first place. Whether that reading survives contact with the mathematics is a separate question, and the framework is deliberately built so that the question can be asked precisely rather than argued about.

Four primitive objects, and nothing else. In the paper's notation: a positive semi-definite correlation kernel K, which says how strongly any two elements are related; a constraint density ρ𝒦; a projection-parameter measure Λ; and a family of projection kernels Π, which smooth the raw correlation structure into something observable. None of these is a spacetime, a coordinate, or a distance. They live on a primitive measurable space with no geometry attached.

On top of the primitives sit five admissibility conditions, labelled A1 to A5. These are the framework's licences: geometric and causal language may be used where they hold, and not otherwise. This is the mechanism by which the paper enforces its own scope — and it is used against the framework as readily as for it. Condition A5, which would license coupling to matter, fails for the instantiation the paper actually works with. That failure is printed in the paper's own list of things it does not claim.

The central technical statement is a conditional reconstruction theorem. Under stated regularity and non-degeneracy hypotheses, and in a suitable coincidence limit, the two-point correlator G₂ defines a metric-like tensor by

In words: take the logarithm of the correlation between two nearby points, differentiate it twice, and what you get behaves like a metric — the object that, in general relativity, tells you distances and angles. Strong correlation reads as short distance. The geometry is not assumed; it is extracted from the correlation structure, which is exactly the reversal the whole programme is about.

Two restrictions are attached immediately, and they matter more than the formula. First, the reconstruction is Euclidean-section only for the instantiation used — it does not by itself deliver the time direction that makes spacetime spacetime. Second, Lorentzian causal language — past, future, light cone, horizon — is licensed only after a declared static or stationary analytic continuation together with a compatible principal-symbol check. Where those are not supplied, causal vocabulary is not used. A framework about spacetime that refuses to say “causal” without a permit is unusual, and deliberate.

The framework's main derived quantity is the spectral dimension, written d_s(ℓ). The idea is measurable rather than philosophical: let something diffuse through the structure and watch how often it returns to where it started. Returns are frequent along a line, rarer on a surface, rarer still in three dimensions — so the return rate encodes the dimension. Because you can watch over a short interval or a long one, the quantity can be computed separately at each scale. Dimension becomes a function rather than a constant. In the paper it is defined from a heat-trace functional.

The framework predicts that in its locked instantiation this quantity undergoes a smooth crossover at a finite calibrated scale. Note the word smooth. An earlier and much more quotable version of this claim said the change was an abrupt step; three pre-registered campaigns in 2026 failed to reproduce any step, and a short proof then showed that the framework's own smoothing operation makes a sharp feature unobservable in principle. The claim was retired, and the retirement is documented separately. What survives is a location, not a jump.

Six entries, each with an envelope, an instrument, a condition for being decidable, and a falsification rule. This is the paper's own register, condensed:

Two of the six are nulls — predictions that nothing happens. That is not padding. A framework that predicts zero somewhere is exposed there: a confirmed non-zero measurement kills it, and a confirmed zero earns it nothing. Registering nulls in advance is one of the cheapest honest things a theory can do, and one of the rarest.

Of the six, exactly one has been confronted with public observation in the paper. The running of the scalar spectral index — roughly, whether the early universe's fluctuation spectrum tilts slightly more at some scales than others — was fitted with a Markov-chain inference against the Planck 2018 likelihoods, using standard cosmological software.

The paper's own reading of this number is the part worth copying. One observable agreeing with a posterior at under one sigma does not constitute confirmation — it is stated in exactly those words, twice. A wide prediction that overlaps a wide measurement is a weak test; both intervals here are broad enough that agreement was fairly likely in advance. It is a consistency check that the framework passed, and the paper declines to inflate it into anything more.

Most papers bury their limitations. This one prints them next to the claims, in the introduction, as a matching list. Reproduced in substance:

General relativity is written in a language whose basic noun is the spacetime point: things happen somewhere, distances are given at the outset, and geometry is what bends. If spacetime is itself an approximation to something more basic — the way temperature approximates the motion of molecules — then no work done inside that vocabulary reaches the level underneath. That is the standing argument for pregeometric approaches, and PCT is one of many: causal sets start from a discrete ordering of events, loop quantum gravity from quanta of area and volume, causal dynamical triangulations from a sum over simplicial geometries, asymptotic safety from a high-energy fixed point, and tensor-network approaches from patterns of entanglement. Several of them independently find the effective dimension dropping at short scales.

So the idea is well populated, and the paper says so. Its own contribution is aimed elsewhere: its motivation is stated as methodological rather than ontological. The complaint, in its own terms, is that across these programmes it is often unclear which observable, measured to what precision, would falsify the framework. Predictions tend to be qualitative — its example is the widely quoted claim that the spectral dimension tends to 2 in the deep ultraviolet — stated without a quantitative envelope. Null tests are rarely specified before the data are examined. The boundary between a framework's legitimate reach and its overreach is seldom drawn. The result is a literature where emergence claims are abundant and contact with observation is sparse.

The response is a framework built to be pinned down. Its inputs are explicit and minimal. The geometric and causal language it derives is licensed only where stated conditions hold, and is otherwise withheld. And a small set of observables is stated in advance, each with a numerical envelope and an explicit statement of what would count as failure. That last part is the point of the exercise: not to be right, but to be checkable.

Set against the programmes listed earlier, this one is unusual in two ways, neither of which is a claim of superiority.

The first is the choice of primitive. Where causal sets start from discrete order, loop quantum gravity from quantised areas and volumes, and tensor networks from entanglement, this framework starts from a correlation kernel — a single object saying how strongly things are related — and derives distance from it. That is a modest-sounding starting point with an aggressive consequence: distance is a summary statistic of correlation, not a container for it.

The second is scope discipline. The framework is explicitly kinematic, and describes itself that way throughout. It is not competing to be the final theory; on its own account it cannot be, since it has neither dynamics nor matter. What it offers is a vocabulary in which emergence claims can be stated precisely enough to be wrong — and the paper's argument is that this is currently the field's scarcer commodity.

Whether that is worth doing is a judgement a referee is currently making. The honest position while waiting is the one the paper takes about its own Planck result: this is compatibility with a research direction, not confirmation of it.