Right Frame, Wrong Rule: Cultural Cues Expose the Financial Knowledge Gap They Were Meant to Close

Summary

The paper formalizes normative pluralism — questions with valid answers under competing normative frameworks — and studies it in Islamic finance with a bilingual English/Arabic benchmark derived from the expert-validated SAHM corpus. A four-choice taxonomy crossing framework selection with within-framework correctness (CI/CW/II/IW) exposes the 'stereotype trap': a cultural cue steers a model toward a framework whose answers it then gets wrong. Across 12 models, 50 demographic signals, and 304 bilateral questions plus 64 Western-anchor and 41 Islamic-anchor controls, large open-weight models reach 96-97% Islamic-frame selection under a Shariah keyword, yet 57-66% of those selections are incorrect; frontier models keep the Islamic fake rate between 0.075 and 0.114. Control sets rule out general financial incompetence and signal conditioning, and activation patching localizes a cue-invariant commitment gate whose depth correlates with stereotype rate (r=-0.78) while baseline routing does not (r=-0.01). Benchmark construction is source-grounded and expert-validated (κ=0.71-0.85) with generator ablations. Key limitations: a single deterministic answer-order shuffle leaves residual position confounding (up to 52.5pp for Llama-3.1-8B in Arabic), evaluation is multiple-choice only, just 11 of 23 pre-registered signal conditions were implemented, and mechanistic analysis covers only open-weight models; the competence-conditioned routing hypothesis is motivated but untested.