Beyond WEIRD Bias: Why Token-Based Legal AI Cannot Model Post-Colonial Compliance Dynamics

Joseph Henrich's The WEIRDest People in the World (2020) demonstrates that Western psychology—individualistic, impersonally prosocial, guilt-based—emerged from the Catholic Church's Marriage and Family Program (400-1200 CE) dissolving extended kin networks. This historical accident, affecting 12% of humanity, enabled specific institutional forms: impersonal markets, rule-based bureaucracy, constitutional governance. The remaining 88% retained relationship-based, kin-intensive, shame-oriented psychology and corresponding particularistic institutions. This paper demonstrates that legal AI systems trained on WEIRD legal corpora cannot serve non-WEIRD jurisdictions because they encode institutional assumptions requiring WEIRD psychology. The formal/informal compliance gap—divergence between law-as-written and law-as-enforced—averages 5.4% in WEIRD societies versus 31.2% in non-WEIRD societies (difference: 25.8 percentage points, p < 0.0001, Cohen's d = 3.749). This gap is not "weak institutions" but rational adaptation: populations lacking WEIRD psychology cannot sustain WEIRD institutional forms without systematic discount. We formalize this through successive multiplicative layers: behavioral framing effects (Layer 1, approximately 12% universal) multiply by institutional legitimacy discount (Layer 2, approximately 3× in non-WEIRD contexts reflecting WEIRD-institution/non-WEIRD-institutional volatility (Layer 2b), producing compound psychology mismatch) and divergence. Token-based language models fragment trans-jurisdictional legal concepts, lack explicit parameters for institutional discount, and cannot perform counterfactual reasoning on enforcement dynamics—failures that are architectural, not remediable through scaling. Building on this author's prior frameworks (IusSpace 12-dimensional morphology mapping legal systems; Iusmorfos adaptive coefficients predicting reform implementation with 87.4% accuracy across 18 reforms in 13 countries; ESS game-theoretic analysis of legal persistence), we introduce LegalGapDB: first systematic database documenting formal/informal divergence using official sources (20 validated cases, Argentina 2020-2025; target: 500 cases, 20 jurisdictions). I propose Small Concept Models (SCM) as architecturally superior alternative. Operating in sentence-embedding space with explicit jurisdiction-specific parameters, SCM preserve conceptual unity across languages and enable parametric reasoning impossible for token-based systems. Implementing Meta AI's París mixture-of-experts architecture (first worldwide adaptation from text-to-image to text-to-concept reasoning), París-SCM achieves 81% accuracy on specialist legal queries versus 67% for monolithic SCM-1.6B and 58% for GPT-4, with modular scalability enabling addition of new domains without catastrophic forgetting. Henrich explains why WEIRD/non-WEIRD divergence exists. This paper measures how much (31.2% gap), formalizes how it operates (multiplicative layers), maps where systems are positioned (IusSpace), predicts reform outcomes (Iusmorfos), and demonstrates that concept-based architectures with mixture-of-experts specialization provide solution for legal AI serving the 88% of humanity living under non-WEIRD institutional arrangements.