A novel transformer-XAI framework with bias attribution and A-RAG for Indian judicial system

Judicial decision-making involves complex reasoning based on legal facts, statutory interpretation, and precedent analysis. Although transformer-based language models have improved the automated analysis of legal documents, their opaque decision-making processes limit their reliability in sensitive judicial applications. The lack of transparency and systematic bias evaluation in existing legal AI systems raises important concerns regarding accountability, fairness, and ethical deployment. This study proposes a transformer-based explainable artificial intelligence framework for judicial verdict prediction using structured representations of legal documents. The framework combines contextual language modeling with attribution-based interpretability techniques to examine the reasoning process of the model. A novel metric, the Bias Attribution Index, is introduced to quantify the influence of sensitive attributes on model predictions. We evaluate NEXAJudicia on two different tasks using data from the Indian judicial system. First, we demonstrate high accuracy in bias classification on a subset of cases annotated for bias (Accuracy: 93%, Macro-F1: 0.91, Weighted-F1: 0.93). Next, we present results on multi-class verdict prediction on 4,001 cases, after a 80–20 split, with fivefold cross-validation (Accuracy: 82.88%, Macro-F1: 0.64, Weighted-F1: 0.89). Furthermore, an attribution analysis shows that the model is driven predominantly by legally relevant evidence from facts and context as opposed to sensitive demographic attributes, which enhances model interpretability and fairness. Summary of NEXAJudicia Paper Abstract We proposed an explainable, bias aware model for legal judgment prediction named as NEXAJudicia which includes hybrid transformer model along with a two stage evaluation process. We used Bias Attribution Index (BAI) for calculating the effect of bias in the model. The observed BAI values remained consistently low across evaluated cases, suggesting limited attributional dependence on predefined sensitive demographic indicators. The prediction uncertainty in legal models is highly dependent on the legal ambiguity and less dependent on the model. Our model predicts with an accuracy of around 81%. The validation of our model was carried out with attribution-aware validation as well as A-RAG-based bias auditing in order to ensure that AI-assisted judicial decisions are interpretable and accountable.