When evolution cheats: Frozen-weights Baselines reveal static-solvers interference in evolved plastic spiking neural networks

In artificial systems, neuroevolution with synaptic plasticity promises agents that adapt during their lifetime, yet standard fitness evaluations do not distinguish inherited static competence from within-lifetime adaptation. Evolutionary algorithms can produce \emph{static solvers}: genomes whose inherited topology performs well even when synaptic plasticity is disabled and weights remain at their lifetime initialization, confounding adaptation with evolved structure. To expose this confound, we introduce a two-phase evaluation protocol with an explicit static-solver extinction mechanism that measures each genome both with plasticity disabled (frozen-weight) and enabled (plastic-weight) on the same task. The first phase quantifies the static competence of the inherited network structure, while the second adds online learning through spike-timing-dependent plasticity (STDP). The \emph{extinction} mechanism removes static solvers from the gene pool, i.e., any genome whose frozen-weight fitness exceeds a threshold, thereby exerting evolutionary pressure for learning and against static solving. We combine this protocol with a complete $2{\times}2$ factorial design that independently toggles extinction pressure and within-lifetime learning. The method is instantiated on a mutable cart-pole benchmark using NEAT-evolved spiking networks with inherited topology and per-connection STDP parameters, while synaptic weights are reinitialized each lifetime and STDP updates run continuously at inference time during evaluation. Across $30$ replicates per condition, the protocol reveals a clear static-solver fingerprint: plasticity-free neuroevolution reaches high training fitness but suffers a roughly $33$-point train--test gap on held-out pole lengths. Learning-enabled agents reduce this gap to about $9$ points and achieve the highest held-out AUC. Extinction does not improve mean test performance when learning is available; instead, it reduces replicate variance and strengthens attribution by preventing high-fitness frozen-weight genomes from dominating selection. These results position extinction not as a generic performance booster, but as a diagnostic intervention for disentangling inherited structure from genuine within-lifetime adaptation.