This paper introduces a new AI architecture: a software compute substrate in which typed, weighted graphs serve as an autonomous computational medium, with properties similar to the human brain yet with full visibility into its inner workings. The premise is simple. Use typed, weighted graph structures to connect labeled neurons at any level of abstraction, then let the graph compute. We show that a single primitive: keyed weight sets with background propagation, unifies eight cognitive capabilities through parameterization: attention (continuous relevance surfacing), imagination (hypothetical injection and diff), simulation (reweighted sensitivity analysis), prediction (forward loop execution through irreducible pathways), intuition (low-threshold peripheral early warning), surprise (prediction-mismatch detection), learning (gated Hebbian updates with hop-count-dependent causal selection recovering STDP), and memory (activation persistence as standing waves under decay). Nodes operate as sovereign agents selecting their own inputs via typed link layers. Under scarcity pressure the substrate self-factors, discovering shared substructure and spawning intermediate abstractions without external direction. Hop count encodes causal proximity: space encodes time, topology encodes causality. The graph compute substrate provides what large language models lack: persistent world modeling, autonomous background computation, and parallel scenario simulation. The two form a symbiotic architecture whose center of gravity shifts as the substrate matures. Empirical validation through production-scale graphs (470+ neurons, 1,400+ connections) demonstrates prediction accuracy improving from 17% to 75% through iterative structural learning. Every activation path is traceable. The system is transparent by construction and irreducible by nature.