Experimental Confirmation of Spontaneous Coherence Condensation in an Octonionic Language Model

The Relational Ontology Framework (ROF) proposes that reality emerges froma non-associative Primordial Relational Field valued in the octonions. A centralprediction is that any octonion-based system, when driven to minimize ontologicalfriction and maximize coherence, will spontaneously condense into a high-coherence,low-friction state a “baryonic node” in the field. We test this prediction us-ing an octonion-valued transformer language model (ROF-LLM) trained on theWikiText-103 corpus with additional loss terms that directly implement the ROFdynamics. A real-valued control model of identical capacity fails to exhibit thecondensation. We further subject the framework to a battery of falsification tests:scrambled octonion multiplication, quaternion control, training on random tokensequences, and loss ablation. All structural attacks fail the true octonion algebra,non-associativity, meaningful data, and both loss terms are individually necessaryfor the observed phase transition. These results provide the first computationalevidence that the octonionic architecture defines a unique attractor not accessi-ble to standard neural networks, confirming the core mathematical claims of theRelational Ontology Framework