Geometry vs. Prediction: Two Theories of What Intelligence Is
The dominant theory of machine intelligence treats reasoning as next-token prediction over a statistical surface. A geometric theory treats it as the preservation of relational structure under transformation. The two are not stylistic preferences — they make different, testable claims about what a system can be trusted to do.
A prediction theory of intelligence says: given a context, produce the most probable continuation. The measure of success is likelihood on held-out data. Under this view a large-enough model, trained on enough sequences, converges toward something we are willing to call reasoning because its outputs become, on average, hard to distinguish from a reasoner's.
A geometric theory of intelligence says something different. It says: given a set of entities and the relationships between them, produce outputs that preserve the shape of those relationshipsA structural preservation property (analogous to a homomorphism in algebra): if A relates to B in the input, A must relate to B in every derived intermediate. Traceable at each derivation step. under any transformation the system is allowed to apply. The measure of success is not likelihood. It is structural fidelity.
The two theories disagree on what a mistake is
Under the prediction theory, a mistake is a low-probability output. It is graded on the surface — the answer sounds wrong. A model can be confidently wrong and there is no internal object to point to as the location of the error.
Under the geometric theory, a mistake is a broken edge. Somewhere in the traversal from premise to conclusion, a relationship that should have been preserved was dropped or fabricated. The location is not statistical. It is topological. You can point to it.
This distinction is not academic. Every domain in which we ask a system to reason about consequences — legal, financial, medical, sovereign, forensic — is a domain in which we need to know **where** the error is, not merely **how likely** it was. A probability score on a whole answer is a diffuse alarm. A broken edge is an address.
Why the prediction theory got so much further
Scale worked. It is a strange and important fact that a system trained only to predict the next token, given enough parameters and enough sequences, produces behavior that looks like reasoning for a wide range of tasks. This is not a rhetorical concession. It is why the field is where it is.
But "looks like reasoning for a wide range of tasks" and "can be trusted to reason correctly on a specific claim" are separated by a gap that no amount of scale has closed. The gap is exactly the one the geometric theory is trying to describe: the difference between a plausible output and a structurally preserved one.
What a geometric system commits to
A geometric intelligence system commits to three things a purely predictive system does not:
**One.** Every claim points to the entities it depends on, and every entity points to the evidence it derives from. The graph is not a decoration on the output — it is the output.
**Two.** Transformations preserve structure by construction, not by loss function. If two claims are related in the input, they remain related in every derived form. This is what "reasoning preserves shape under transformation" means concretely.
**Three.** Verification is a walk along edgesThe verifier does not re-execute the reasoner. It traverses the same edge set the reasoner claims to have walked, and checks that every edge exists in the evidence graph., not a re-derivation from scratch. You do not re-run a model to check an answer. You trace the edges the answer walked.
Neither theory is finished
The geometric theory is not a complete replacement. It is a claim about the correct primitive. Whether the primitive is a graph, a manifold, a categorical structure, or something not yet named, it is not a token distribution.
The rest of this series works out what follows from taking that seriously.