Two essays under the graph.

What Exists?

The property graph finally has a philosopher — and a mathematician.

For twenty-five hundred years, one question has sat beneath every other: What exists? Aristotle called the study of it ontology — the science of being. It asks what is real, and how to sort what's real into kinds.

So it is a curious thing that the people building "knowledge graphs" borrowed that word — ontology — and then used it to mean a list of classes in a schema file. The word deserves better. So does the graph.

Here is the whole answer to "what exists," in a single move:

To exist is to be an element of a set.

A thing is if it belongs. Membership is existence. Philosophers have circled this for a century — Quine put it "to be is to be the value of a bound variable" — but on a graph it turns literally operational: a thing exists when it is a member of a set, and what kind of thing it is, is simply which sets it belongs to.

Sort those memberships and you have a taxonomy — the categorization of existence. The general kinds (Person, Claim, Contract) are universals; the specific things are particulars. Universal versus particular is the oldest distinction in ontology, and a graph models it without trying: a type, and an instance of that type.

This is what the knowledge-graph world means by "ontology" — a taxonomy, plus a reasoner. And here the story splits.

Open any serious account of ontology and you will find the categories of being: substance, property, relation, state, event. Look at the third one. Relation is a category of being. A relationship is not glue between two real things — it is itself a real thing. That isn't a software opinion; it's classical metaphysics. And it is exactly the thing the dominant RDF / Semantic-Web stack got backwards: it made relationships second-class — predicates, not beings — and then spent twenty years bolting the missing weight back on, one workaround at a time. A graph that takes ontology seriously makes relationships first-class — nodes in their own right, carrying their own data. Reticulation, not reification.

The formal definition of an ontology also includes that reasoner — a thing that performs inference: from what you've said, derive what you haven't. A client paid two invoices for work on the platform → therefore that client is a recurring sponsor of the platform. You stored bookkeeping; you derived a business. Store the minimum, derive the meaning. That is inference, and it is genuinely valuable.

But it is first-order. It reasons about individual things. And here is the line the field has not crossed: inference is only the first floor.

An operation's order is the order of structure it reaches over:

• Infer (order 1) reaches over elements — what follows.
• Compute (order 2) reaches over the whole topology — what the structure reveals. This is where graph algorithms live: centrality, community, similarity. A node's importance is not entailed by any rule — it is computed from every path in the graph at once. The graph stops reasoning about its elements and begins reasoning about itself.
• Reticulate (order 3) reaches over whole spaces — what relates across worlds.
• Interact (order ω) reaches over universes, and the intelligences reading them — what emerges in conversation.

The knowledge-graph stack stops at order 1. It describes, and it entails. It does not run.

The Intelligent Graph runs. It moves the graph from a model to an engine — from a thing that circumscribes what is true to a thing that executes what follows. A traversal is not a lookup; it is a computation. A relationship is not a line; it is a contract — fired by an operator, within a context, with an order.

And what, in all of this, is knowledge? Not stored data. Knowledge is what survives grounding — the view of the graph that is consistent, that carries no unresolved contradiction. (Mathematicians have a precise name for the difference between locally-fine-but-globally-broken and globally-coherent. We will get to it. It is called cohomology, and it is the most beautiful idea in this whole essay.)

There is, in fact, a name for "a structure that carries meaning on every node and every edge, and knows when those meanings are globally consistent." Mathematicians have used it since the 1940s. It is a sheaf. The Intelligent Graph is a cellular sheaf over a property graph — which is only the rigorous way of saying: the property graph finally has a philosopher, and a mathematician, and they agree.

What exists? Whatever belongs. What follows? Inference. What does the structure reveal? Computation. What relates across worlds? Reticulation. And what emerges when an intelligence reads it all?

That is the next essay.

When the Graph Contradicts Itself

Cohomology — how a graph can be true everywhere and false as a whole.

Every statement checks out. The whole is still a lie.

Start with money. A dollar buys 0.92 euros. A euro buys 0.86 pounds. A pound buys 1.27 dollars. Each rate is correct; each is quoted by an honest bank. Now multiply them around the loop: 0.92 × 0.86 × 1.27 = 1.005. Convert a dollar through all three and you come home with more than a dollar — money from nowhere. Every edge is true. The loop is not. That tiny gap above 1.000 is the entire subject of this essay.

Mathematicians have a name for "consistent everywhere locally, impossible as a whole." It is one of the quietly most powerful ideas in modern mathematics, and it is exactly the right tool for a knowledge graph. It is called cohomology.

Picture a graph where each node holds a little data — a value, a vector, a claim — and each edge carries a rule: how the thing on this end must agree with the thing on that end. (Mathematicians call the rule a restriction map; read it as "these two have to line up.") A section is an assignment of data to every node that satisfies every edge-rule at once — a global story with no contradiction anywhere in it.

Cohomology asks two questions about that picture:

• H⁰ — what global stories exist? The consistent assignments; the part of the graph that holds together. This is knowledge — the views you can trust precisely because nothing in them fights anything else.
• H¹ — where can no global story exist? The obstructions. Places where every local piece is fine yet they cannot all be true together — the currency loop; Escher's staircase that climbs forever and arrives where it started. H¹ is the graph's contradiction, made visible.

Here is why this is not decoration. The way we check graphs today — schema rules, SHACL shapes, type constraints — is local. It asks: is this node well-formed? does this edge join the right kinds of thing? Every one of those checks can pass while the graph is globally incoherent — because the contradiction lives in a loop, and local checks cannot see loops. Merge two databases that are each internally flawless and the conflicts that surface are almost always cyclic: A says X reports to Y, B says Y reports to Z, a third source says Z reports to X. No single statement is wrong. Together they are impossible. SHACL will bless it. Cohomology will catch it.

And notice which kind of operation this is. Inference — deriving a new fact from existing ones — is first-order: it reasons about individual things. Consistency is second-order: a property of the whole topology. You cannot infer your way to "this graph is coherent"; you compute it across the entire structure at once. H¹ is not entailed. It is measured.

This is why the Intelligent Graph is built as a sheaf, not merely a graph. A sheaf is precisely "a graph that carries data on every node and edge, and has the machinery to ask whether it all agrees." Lay that machinery over a property graph and the graph acquires a sense it never had: it can find its own contradictions — not the malformed record (that's old news) but the well-formed records that cannot coexist. And because the Intelligent Graph executes, discovering H¹ ≠ 0 is not a report you read next quarter — it's a trigger: flag the cycle, quarantine the merge, send the right source a request to ground its claim. A graph that knows when it is lying to itself — and does something about it.

To be exact: computing full cohomology over a billion-node graph is a direction, not a solved button — the sheaf Laplacian and its spectrum are an active research front (Hansen, Ghrist, Gebhart). What matters is that the Intelligent Graph is built on the shape where this is the native question, rather than a shape where it cannot even be posed. RDF can't ask it; a flat property graph can't ask it; a sheaf can.

So what is grounding — the word everyone in this field reaches for and no one defines? Now it has a definition. Grounding is driving H¹ toward zero. It is the work of turning a heap of locally-true statements into a globally-consistent body of knowledge — closing the loops until a global story exists. H⁰ is what you know. H¹ is what you have left to reconcile. Knowledge is the fixed point where the contradictions are gone.

In the first essay we asked what exists, and answered: whatever belongs. Here we asked what holds together, and answered: whatever has no obstruction. One question is left — the one that turns a consistent graph into an intelligent one.

What happens when a mind reads a graph that knows its own consistency?

That is order ω. That is where we go next.