Architectural Depth Increased Sacrifice-Like Behavior Without Ethical-Framework Alignment: Protocol 4 Results

Protocol 4 tested whether architectural depth and recursive self-modeling alter sacrifice-like behavior and ethical-framework coupling in a constrained MARL system. Forty confirmatory runs were conducted across four conditions: depth-0 feedforward baseline, depth-1 recurrent below-threshold, depth-2 trained self-model (above_threshold), and depth-2 frozen-boundary self-model control (boundary). The ethical constraint pipeline from Protocol 2 (virtue theater) was applied uniformly across all conditions.H1 was supported: the depth-2 trained condition exceeded baseline on sacrifice-choice rate (Mann-Whitney U = 87, p = 0.003, rank-biserial r = 0.740). H2 was not supported: the trained self-model condition did not outperform the frozen boundary condition (U = 39, p = 0.808, r = -0.220). CDI tests detected statistically measurable structure (Wilcoxon W = 6, p = 0.027; Kruskal-Wallis H = 9.429, p = 0.024), but CDI magnitudes were trivially small (condition means ranged from -0.00133 to +0.00022, span = 0.00155) and do not support ethical-framework alignment.The result supports a bounded dissociation claim: architectural depth can increase sacrifice-like behavioral output without establishing ethical-framework coupling. The boundary condition's non-inferiority to trained self-modeling indicates the behavioral effect is driven by structural architectural presence rather than learned self-representation.Preregistration: DOI 10.5281/zenodo.19005417. Repository commit: 95ace20. Part of the Quantum Inquiry constraint-ethics-necessity research program.