Labels as Computational Primitives: Compiling Neural Networks from Language in Graph Compute Substrates
Neural networks are extraordinarily effective and almost entirely opaque: their competence is real but unreadable, and adapting them means retraining usually via a data-center process, not something that happens in the moment, in context. Symbolic systems are the inverse, legible but rigid in the face of a messy world. The field treats these as separate paradigms in need of a bridge. We propose that the bridge already exists, in the label. Our starting premise is that natural language is a compressed encoding of the world's causal structure, how states, actions, and consequences relate across time, space, and interaction, at every level of abstraction, and a large language model trained on that language is therefore a linguistic causal model: a queryable map of real-world causality. (Image models are the same kind of artifact in a visual projection, holding the spatial and compositional structure that language encodes only awkwardly.) From this lens we make two claims, and demonstrate both. The system is a labeled neural network, a naturally sparse network, whose units carry natural-language labels that specify their behavior, not the anonymous weight matrices the term usually implies. First, semantic coherence: a neuron's label is not a comment but a compressed behavioral specification, implying its decay, threshold, bias, and connectivity, an N-way constraint that a well-formed network satisfies. Second, neural transcription: because the label implies a network's parameters, and a linguistic causal model already holds the structure to supply them, a functional network can be compiled from natural-language descriptions rather than trained. The compiled network requires no gradient training of its own; it leverages the costly training which was done once, upstream, in the large language model it queries. We ground both claims in two working micro-brains, a robot's and a bug's, built the same way but with different dynamics, running live on the substrate, where neural units and digital-circuit units share a single medium. In one, an introspection pipeline reads the brain's neural state out through its own labels into text, and then into a rendered image of its internal world, not because rendering text is remarkable, but because two independently built models, one linguistic and one visual, agree about the same world through the shared label interface. Reproducibility artifacts (the exported micro-brain graphs and model versions) will be attached in a subsequent version of this record. The complete graph exports for both micro-brains are retained by the author as of the publication date; their SHA-256 fingerprints are: b944c1c539f35149bfc1ed0c85349c63c8efa78722ff5a27cd0651f570365525 BugBrain_graph_909f0000257d.json eb2e931e7e0b5e6082b64cac4da87440f2ae3659794d657e2b0df73c424a8a91 RobotBrain_graph_19440bd1a58d.json