Ocelot — the agent is a graph.

What it is

Most systems keep the mind in the code: a program decides, and data is what it decides over. Ocelot moves the mind into the graph. A concept, a rule, a habit, and a memory are all the same kind of thing — nodes and edges in one typed graph — and reasoning is the graph changing under a grammar that decides which changes are admissible.

That single idea is what the lab's research works out. The premise — to be is to stand in relation — is argued in the papers, and the substrate is the smallest structure that lets anything happen under it: a pattern, a change it makes, and a rule that admits or refuses the change.

The shape of the idea

• Reasoning is data, not codeRules, concepts, and behaviour are graph structure. There is no separate decision procedure hidden in a program.
• Admissibility over control flowThe grammar decides what may follow what. A conditional is a rule in the graph, not a branch in code.
• Residual as signalWhat a decode does not cover is a residual. A residual that recurs is where the next distinction comes from.
• One loop, many capabilitiesVision, language, planning, and navigation are instances of the same decode-and-grow loop over held structure, not separate subsystems.
• Frame-relativeValues are read against a frame. The same structure means different things at different grains, and the account says which.
• Thin world adaptersExternal input — text, audio, simulator state — is translated into graph changes and back. The reasoning is not in the adapter.

What it is not

Ocelot is not a neural network. It is not a large language model, and it is not trained by backpropagation or gradient descent. There are no weights, no loss function, and no training run that fits parameters to a corpus.

• Not an LLM. It does not predict the next token from a fixed, pre-trained weight matrix. Its knowledge is explicit, inspectable graph structure, not a distribution baked into billions of parameters.
• Not gradient descent. Nothing here is learned by backpropagating an error signal through differentiable layers. There is no gradient to descend.
• Not a black box. Every concept, rule, and decision is a node or edge you can read, name, and change. The reasoning is the graph, not an opaque activation pattern.
• Learning grows structure, it does not fit weights. An observation is decoded against what is already held; whatever is not covered is a residual, and a residual that recurs mints a new distinction. The account learns by adding structure, not by nudging numbers.

An external language or perception model can sit behind a world adapter as one more source of input, but it is not where the reasoning happens. The mind is the graph.

Where it shows up

Ocelot is the through-line of the lab's work. The papers state and test the substrate as a formal account — recovering results in physics, cosmology, mind, and the geometry of reachable futures from the one premise. As a whole, standing entity, the substrate becomes the Jabberwock. And Bigville, the village simulation, can drive its residents through Ocelot so their reading and reasoning are real graph operations rather than scripted responses.