This paper introduces CL-SHIFt, a framework enabling Sharia chatbots to adapt to evolving jurisprudential consensus and user needs through real-time user feedback, while maintaining doctrinal consistency and preventing forgetting of canonical knowledge.
Key findings
Proposes a Dual-Memory Architecture to separate sacred knowledge from adaptable jurisprudential interpretations.
Introduces a Feedback-to-Reward Translation Module to convert user signals into optimization targets.
Includes a Madhhab-Aware Constraint Layer to ensure updates respect school-specific legal methodologies.
Establishes evaluation benchmarks for doctrinal consistency, cross-madhhab generalization, and expert validation.
Limitations & open questions
The framework's effectiveness in diverse real-world scenarios needs further validation.
The integration of human feedback may introduce biases that require careful management.