Hybrid QLoRA-RAG Architecture for Saudi End-of-Service Benefits Calculation: Synthetic Data Generation and Uncertainty Quantification for Legal Reasoning

Deploying large language models for high-stakes domain-specific reasoning requires addressing challengesabsent from standard benchmarks: handling incomplete information, quantifying uncertainty, and performing multi-step numerical calculations with authoritative source attribution. We present a hybrid architecture combining parameter-efficient fine-tuning via Quantized Low-Rank Adaptation (QLoRA) with Retrieval-Augmented Generation (RAG), evaluated on Saudi Arabia’s End-of-Service Benefits calculation—a legally binding financial computation involving 16 interacting legal provisions across 35 termination scenarios. Our contributions include: a comprehensive synthetic dataset of 10,000 samples systematically modeling real-world legal consultation complexities—incomplete information (15%), conflicting evidence (10%), legal interpretation ambiguities (5%), and adversarial examples (5%)—grounded in empirical distributions from 47,382 actual cases, 3,847 labor court disputes, and expert interviews (n=23); a hybrid architectural approach demonstrating that combining QLoRA fine-tuning (0.42% trainable parameters, 93.5% memory reduction) with retrieval-augmented generation yields complementary benefits, outperforming isolated components by 5.8–8.7 percentage points;and integrated uncertainty quantification mechanisms combining epistemic (MC Dropout), aleatoric (retrieval confidence, linguistic hedging), and calibration (temperature scaling) methods achieving Expected Calibration Error of 0.043 and 89.4% precision in detecting ambiguous cases requiring human review. Evaluation on 1,000 held-out synthetic test cases—stratified across six complexity tiers—shows 94.2% accuracy (±5% tolerance), 91.5% legal citation correctness, and graceful degradation across complexity tiers (98.7% standard cases → 82.0% adversarial examples). We note that all quantitative evaluation is conducted on synthetic data; real-world deployment validation remains an important next step. Human evaluation by five Saudi legal experts (inter-rater κ = 0.73) yields 4.4/5 overall rating with unanimous recommendation for pilot deployment. While our primary evaluation relies on synthetic data and focuses on a single legal calculation domain, the methodological framework—synthetic modeling of domain ambiguity, architectural patterns for parametric-retrieval integration, and uncertainty-aware human-AI collaboration—provides a transferable template for specialized reasoning tasks requiring numerical precision, source attribution, and confidence calibration. We discuss threats to external validity and outline concrete steps toward real-world validation.