The Negative Space

I.The First Known Thing — the definition made only of denials

theology · the via negativa, Pseudo-Dionysius

Here is a claim theologians have argued over for fifteen centuries. Writers in the apophatic tradition — Pseudo-Dionysius foremost among them — held that no positive statement about the absolute can be true. Not because we lack the information, but because predication itself is the problem. To say God is good imports a human measure of goodness and draws a fence around what was supposed to be unbounded. Every positive attribute diminishes its subject by the act of applying it.

every positive attribute is a fence

What the tradition offers instead is relentless negation: not this, not that, not the other — and then, for rigour, the negation of the negations themselves. What is arrived at is not a description at all. It is a position reached entirely by exclusion, the way a sculptor arrives at a figure by removing stone. This is established intellectual history, worked over in philosophy as much as in theology.

II.The Second Known Thing — the space too large to point at

machine learning · latent spaces and contrastive loss

Now walk to a different hall of the library entirely. Modern models represent meaning as vectors in spaces of many thousands of dimensions, and such spaces behave nothing like the three we live in. Pick two directions at random in a thousand dimensions and they will almost certainly be close to perpendicular. Volume concentrates in unintuitive places. Distances flatten out until near and far stop meaning much.

the cluster is a speck in an enormous volume

Practitioners meet the consequence daily. Pinning down a concept by pulling its positive examples together is not enough — the resulting cluster occupies a vanishing fraction of the available volume, and under the curse of dimensionality it fails to hold. What works is contrast: negative sampling and contrastive objectives, which push away everything the concept is not and let the region be defined by its boundaries. This is standard, well-established machinery.

III.The Bridge

nobody's textbook contains both chapters

Medievalists do not read papers on contrastive loss. Machine learning researchers do not read Pseudo-Dionysius. Each fact above is old news in its own hall. But hold them in the same frame, and a shared mechanism appears:

X: high-dimensional information bounding

Call it high-dimensional information bounding. When the space of possible states grows exponentially, positive coordinates address almost nothing: any particular point is an infinitesimal share of the volume, and naming it conveys far less than it appears to. What can actually be specified is an intersection of exclusions — a region fenced by half-spaces, each one saying only not on that side. The mystic's endless denials and the contrastive loss function are then the same manoeuvre performed on the same difficulty: an entity too large for positive predication, located by the accumulated weight of everything it is not.

IV.The Question That Becomes Possible

neither discipline would ask this alone

Here is the part that makes this an intercognix candidate rather than a nice metaphor. If the apophatic writers were right that a positive predicate does not merely fail to capture its subject but actively bounds and diminishes it, then the same charge can be laid against a training objective. A positive centroid is a positive predicate: it commits the representation to a location. And a question appears that neither field would ever ask on its own:

the apophatic embedding hypothesis

Does meaning in neural embeddings degrade because positive attribution vectors inherently introduce lower-dimensional compression artefacts — and would a model trained solely on negated contrastive half-spaces show better out-of-distribution reasoning?

why it would matter

If it would, the consequence lands squarely on how representations are trained at all. Out-of-distribution reasoning is precisely where current models are weakest, and the objective function is one of the few levers that reaches every part of a model at once.

V.How It Could Be Proven — or Broken

a candidate must be breakable, or it is merely poetry

An intercognix candidate earns nothing until it names its own test. This one does:

the honest status

Until such work is done, this remains exactly what its stamp says: a candidate. A surprising structural echo between two halls of the library, waiting for someone with the tools of both to walk the corridor.