Does Recursive Model Succession Change Decision Reproducibility by a Simple Rule? A Pre-Specified 600-Cell Lineage Study

This study investigates whether recursive generation of neural-network models produces systematic changes in behavioral decision reproducibility across successive generations. We evaluate a closed 600-cell experimental grid comprising six conditions, five model seeds, five search seeds, and four generations, using small image-classification multilayer perceptrons trained on MNIST. The study uses behavioral fidelity to a fixed surrogate as its primary outcome. The experimental design separates the effects of weight inheritance and proposal mechanism through a 2×2 ReLU factorial, with activation-matched tanh and blind-random control conditions. The analysis evaluates generation trends, adjacent-generation transitions, control comparisons, and a pre-specified multiple-testing family. The primary result is negative or mixed at the level of this behavioral proxy. Pooled generation-level analysis estimates a small negative trend in fidelity (coefficient −0.000868, SE 0.000389, p = 0.0258, 95% CI [−0.001631, −0.000105]), while condition-specific and transition-level results show heterogeneous patterns rather than uniform degradation across recursive generations. The study is presented as an empirical lineage study of a behavioral proxy, not as a new interpretability method or a claim about general AI self-improvement or capability growth.