Selection navigates a degenerate circuit space: behavioral individuation without structural differentiation in constrained neuroevolution

Abstract

Neural degeneracy, the capacity of structurally distinct circuits to perform the same function, is typically studied as an emergent property of biological neural systems. Here we impose it by construction, engineering the conditions that make degeneracy the expected outcome of evolutionary search, and ask what individual-target selection achieves within that degenerate space. We use constrained neuroevolution to evolve 14-neuron recurrent circuits (Dale's Law, sparse connectivity, quantized weights) replicating the natural navigation behavior of 9 individual mice across 54 independent evolutionary runs. Architectural constraints impose a structural floor: no aggregate circuit statistic differs across mice (0/18 features, all pFDR > 0.47), and this uniformity extends to the topology axis itself: topology distance predicts behavioral distance at no scale, in evolved or random constrained agents alike, and no structural axis carries significant information about behavioral identity (maximum NMI = 0.2173). Yet behavioral individuation is robust: own-mouse fitness error is 33.4% lower than cross-mouse error, and this specialization persists on held-out data. We show that what individual-target selection shapes is not circuit structure but the strength of functional sensitivity commitment: specialists develop roughly 6.4-6.7x higher sensitivity variance than generalists trained on all mice simultaneously (9/9 mice, mouse-level p = 0.002), despite exploring statistically indistinguishable topological diversity. The particular pathways a circuit commits to, carry no shared mouse-specific signature, so behavioral individuation is expressed as a magnitude of functional commitment rather than a structural or pathway fingerprint - degeneracy that operates not just at the structural level but at the level of the computational strategies that implement individual behavior.