Intelligence
has a shape.

Today's leading AI systems learn once, then freeze. They cannot integrate new experience without retraining, they fabricate when pressed past their training distribution, and their compute costs grow linearly with everything they have to know.

We believe these are not three independent problems. They are symptoms of a single architectural choice: representing knowledge in weights rather than in geometry. A mind built from a frozen weight matrix will always be one step behind the world it is trying to understand. A different structure is needed for the next generation of intelligence.

The brain already organizes what it knows as geometry. We did not invent this. We are building toward it.

Open a cortex and you will not find a list of weighted parameters. You will find maps. The visual cortex is arranged so that adjacent neurons see adjacent regions of the visual field. The motor cortex is laid out so that nearby cells control nearby muscles. The hippocampus contains place cells and grid cells that form a literal coordinate system for the space an animal is moving through.

Modern neuroscience has a name for the more general phenomenon: neural manifolds. The activity of large populations of neurons traces out continuous geometric structures, and those structures, not the individual cell firing rates, are what carry the meaning. The brain knows things by being shaped a certain way.

We believe an artificial mind worth the name will work on a similar substrate. PCM is built on this premise. The architecture is not a metaphor borrowed loosely from biology. It is the operational form of the same principle: that knowledge is geometry.

PCM is built to do four things conventional AI cannot. These are the operational properties of the architecture, not its mechanisms.

The mechanisms that produce these properties are disclosed under non-disclosure, and the underlying mathematics under a formal evaluation agreement. The properties themselves are how we measure whether the architecture is doing its job.

One intelligence, many interfaces.

Large language models are remarkable at language. World models are remarkable at perception. We believe neither is, on its own, a mind. PCM acts as the integrating layer between them. It queries them when their strengths apply, and absorbs what they return into its own evolving geometry.

A geometric architecture is in operation today, under restricted access, with a roadmap that moves from static manifolds toward federations of cognitive agents that share what they learn without surrendering what they know.