Urdu News Credibility Analysis via Retrieval-Augmented Multilingual Classification: An Initial Baseline and Empirical Study

Automated credibility assessment of Urdu news remains largely unsolved. Most existing systems rely on black-box classification that produces verdicts without retrieving or grounding decisions in external evidence. This paper presents the Urdu News Credibility Analyzer, an initial retrieval-augmented pipeline combining FAISS-based dense evidence retrieval with fine-tuned multilingual BERT (mBERT). Building on limitations identified in a prior English-only News Authenticity Analyzer, we fine-tune mBERT on a cleaned, balanced subset of the Ax-to-Grind Urdu benchmark (5,708 samples) and construct an external Urdu evidence corpus of 321 verified articles from BBC Urdu, Jang, and Express News. Our central finding is a 17.51 percentage point accuracy drop when naive claim-evidence concatenation is applied: mBERT without retrieval achieves 92.12%, while RAG-enhanced mBERT achieves only 74.61%. Confusion matrix analysis reveals this degradation is asymmetric TRUE news misclassification increases from 20 to 117 cases demonstrating that naive evidence fusion biases the classifier rather than grounding it. This empirical result provides a reproducible, mechanistically explained motivation for structured Chain-of-Thought reasoning, reliability-aware evidence ranking, LoRA fine-tuning, and faithfulness evaluation under the planned TruthLens multilingual framework targeting Urdu and Japanese.