Metric aggregation divergence is a hidden confound in agent-based model policy search: when optimization, tournament evaluation, and statistical validation independently implement outcome metric extraction, champion selection reflects aggregation artifact rather than policy quality. We propose Hierarchical Evolutionary Agent Simulation (HEAS), a composable framework that eliminates this confound through a runtime-enforceable metric contract - a uniform metrics_episode() callable shared identically by all pipeline stages. Removing the confound yields robust champion selection: in a controlled experiment (n=30), HEAS reduces rank reversals by 50% relative to ad-hoc aggregation; the HEAS champion wins all 32 held-out ecological scenarios - a null-safety result that would be uninterpretable under aggregation divergence. The contract additionally reduces coupling code by 97% (160 to 5 lines) relative to Mesa 3.3.1. Three case studies validate composability across ecological, enterprise, and mean-field ordinary differential equation dynamics.

翻译：指标聚合分歧是基于智能体模型政策搜索中隐藏的混杂因素：当优化、锦标赛评估与统计验证各自独立实现结果指标提取时，冠军选择反映的是聚合伪像而非政策质量。我们提出分层进化智能体仿真（HEAS）——一种通过运行时强制执行的指标契约（即所有流水线阶段共享的同一套 metrics_episode() 可调用接口）消除该混杂因素的可组合框架。消除混杂因素可实现鲁棒的冠军选择：在对照实验（n=30）中，HEAS相较于临时聚合方法将排名逆转减少50%；HEAS冠军在所有32项保留生态场景中获胜——这一零假设安全结果在聚合分歧下将无法解释。该契约相比Mesa 3.3.1还将耦合代码减少97%（从160行降至5行）。三项案例研究验证了其跨生态体系、企业体系与平均场常微分方程动力学的可组合性。