Synthetic Minds in Normative Sandboxes: Exaptation, Spandrels, and Multilevel Selection in Agent-Based Legal Simulation

This paper introduces the Synthetic Minds Normative Sandbox (SNMS), a second-generation agent-based simulation framework for modeling the emergence and transformation of legal norms within multi-agent institutional environments. SNMS advances beyond first-generation normative automata by equipping agents with cognitively differentiated MindProfiles — composite structures encoding cognitive, affective, and doctrinal dimensions — calibrated empirically to eight Argentine institutional archetypes. Central to the architecture is the integration of TribeV2 (King, Benchetrit, Rapin, Brooks, Begany, Raugel & Banville, 2026), a tri-modal foundation model of human brain activity trained on over 1,000 hours of fMRI data across 720 subjects. In SNMS, the cortical encoding vectors produced by TribeV2 over normative text excerpts serve as a proxy for differential cognitive load across archetypes. The theoretical framework draws on three pillars: (1) Gould & Vrba's (1982) exaptation; (2) Gould & Lewontin's (1979) spandrels; and (3) Wilson & Sober's (1994) multilevel selection. Four novel metrics operationalize the framework: the Functional Drift Index (FDI), the Spandrel Emergence Metric (SEM), the Multilevel Selection Ratio (MSR), and the Mean Compliance Index (MCI). A proof of concept applied to a simulated Argentine labor reform scenario across 100 agents, 20 rounds, and 5 random seeds yields a Compliance Legitimacy Score (CLS) of 0.89, an FDI of 0.42, two normative spandrels by round 14, and an MSR of 0.73.