Cross-Domain Transfer Without Gradient Descent: Emergent Symbol Grounding in a Continuously Running Neural System

We report the emergence of cross-domain transfer in a continuously running, self-organizing neural system that uses no gradient descent, no backpropagation, no attention mechanisms, and no explicit training objective. In controlled experiments, we demonstrate three findings: (1) A meta-cognitive layer correctly detects mastery and cannot be fooled by descriptions of competence. (2) Exposure to natural language text—whether correct or incorrect—provides a ~12–15pp boost on arithmetic, demonstrating structural priming with zero semantic comprehension. (3) When language and arithmetic are temporally interleaved, the system trades memorization for generalization: rote recall drops 12.5pp but performance on never-seen combinations improves 3×. These results emerge from local Hebbian-type plasticity on a reaction-diffusion substrate with no optimizer, no loss function, and no backpropagation.