Architectural assessments of traffic anomaly detectors can be sensitive to data partitioning, random training seeds, and dependence among overlapping windows. We evaluate nine encoder configurations, four fusion operators, and two graph constructions using physics-informed synthetic trajectories generated on an OpenStreetMap-derived road network. Each fold–seed threshold is selected on validation data and applied to its corresponding test set. Trip-level cluster bootstrap and permutation tests account for dependence among windows, while corrected resampled tests approximate the dependence induced by overlapping training sets. On a fixed partition of 14 test trips, no comparison establishes superiority after Holm correction. Across five-fold cross-validation with five seeds, mean F1 scores range from 0.8558 to 0.8839 across fusion operators, with gating and cross-attention producing occasional degenerate runs. Across 500 runs covering repeated grouped partitions and both graph constructions, no pairwise architectural comparison remains significant after correction. With run-specific thresholds, subtype analyses do not support the abnormal-dwell improvement that appears under pooled thresholds. Classical baselines remain competitive, but their apparent advantage in single-seed pooled tests is not corroborated by run-level inference. Overall, the results demonstrate that architectural conclusions are sensitive to threshold selection, the number of training seeds, and the choice of statistical unit of analysis.

In higher education, hallucination-free information retrieval is critical for navigating complex institutional legislation. This study develops a highly accurate, production-ready Retrieval-Augmented Generation (RAG) architecture tailored to Turkish, a morphologically rich, low-resource language. We evaluate the pipeline using an expert-curated dataset of 846 canonical question–answer pairs derived from real-world student inquiries. Rigorous statistical testing reveals that while top embedding models achieve comparable, statistically indistinguishable retrieval accuracy, BAAI/bge-m3 demonstrates robust representational symmetry by maintaining identical document rankings across normalized distance metrics. Challenging common literature assumptions, our architectural ablation shows that dense retrieval alone is highly sufficient; hybrid (dense + sparse) fusion yields no meaningful accuracy gain when indices are constructed over equivalent text. At the generation layer, an efficiently sized 9-billion-parameter model matches or surpasses the response quality of much larger models—a ranking confirmed via repeated-run significance testing—while operating 3.9 times faster than a 32-billion-parameter counterpart and 2.5 times faster than an 8-billion-parameter alternative. Finally, we demonstrate that the latency benefits of INT8 Scalar Quantization are highly scale-dependent, accelerating retrieval in small-scale micro-benchmarks but increasing end-to-end latency at larger index scales. Ultimately, this study establishes that optimizing a production RAG system relies on selecting highly aligned, efficient components rather than maximizing architectural complexity or parameter count.