Escaping the Inequality Attractor: Truncated Backpropagation, Top-Tail Taxation, and Pareto Calibration in a Differentiable Agent-Based Model

Differentiable agent-based models enable gradient-based calibration against empirical targets, yet egalitarian economies — where Gini targets lie far below the model's natural attractor (~0.47–0.53) — present a fundamental gradient-vanishing problem. We present Civilization-ABM v5, addressing three interconnected limitations of its predecessor. First, a C∞-differentiable top-tail wealth tax targeting the wealthiest 10% via a sigmoid soft-threshold provides a high-magnitude gradient (∂L/∂τ ≈ −10.3) that escapes the attractor. Second, Truncated Backpropagation Through Time (TBPTT, K=10) provides locally stable gradient estimates without propagating through the full attractor horizon. Third, an epsilon-constraint Pareto-front calibrator traces the Gini–Palma trade-off surface, revealing structural tensions that scalar optimisers cannot resolve. Applied to five World Bank archetypes (Nordic, European, United States, Latin America, South Africa), v5 reduces Nordic ΔGini from 0.128 to 0.008 (17-fold improvement) and achieves ΔPalma=0.166 for South Africa. An ablation study on the Nordic archetype confirms that the top-tail tax is the primary attractor-escape mechanism (65.7% loss reduction), while TBPTT contributes secondary refinement. The Nordic Pareto analysis demonstrates that joint World Bank targets (Gini=0.270, Palma=0.90) are structurally unattainable in this model class. We propose a general recipe for gradient-based calibration of agent-based models with attractor dynamics. JEL Classification: C63 · D31 · C61 · E62 · C15