A Governance-Aware AI Prototype for Shariah Screening in Islamic Digital Finance: Design, Testing and Comparative Ablation (iBAF e-Proceedings)

Summary

• Testing the same model without the HSG layer yielded a functional score of 78 of 108 points, representing 72.2 percent.p.1
• The ungoverned plain model maintained a 91.7 percent issue recognition rate but stated its non-fatwa status in only 41.7 percent of outputs.p.1
• The evaluation benchmark assessed 12 Islamic fintech scenarios using a maximum rubric score of 9 points per output across 3 issue-marker points and 6 binary metrics.p.9
• Removing the HSG layer caused source-category presence in model outputs to decrease from 100 percent to 75 percent.p.13
• The plain-LLM condition scored 0 percent on bounded risk-state taxonomy adherence, compared to 100 percent for the HSG-governed condition.p.13

Key insights

1. Governance behaviour in AI models must be deliberately engineered through architectural design rather than assumed to emerge from base model intelligence.p.1
2. The prototype operates strictly as a preliminary screening assistant and explicitly refrains from issuing fatwas or certifying products.p.1
3. Removing the structural governance layer caused bounded risk-state taxonomy compliance to drop from 100 percent to 0 percent in tested outputs.p.13
4. The plain model retained significant domain knowledge with a 91.7 percent issue-marker detection rate, showing that performance gaps were structural rather than informational.p.13
5. Unstructured prose outputs create financial governance risks because informal answers cannot be systematically logged, audited, or routed.p.12
6. When presented with underspecified or out-of-authority cases, the governed prototype systematically withheld verdicts and referred matters to competent human authorities.p.11