Many judicial systems face severe backlogs, particularly in countries with limited judicial capacity such as India. Recent advances in Large Language Models (LLMs) offer new opportunities to support legal workflows, yet their reasoning remains opaque and difficult to evaluate. Existing datasets and benchmarks for legal judgment prediction (e.g., CAIL, ILDC, LexGLUE) primarily treat the task as classification, focusing on outcomes rather than the structure or quality of legal reasoning. This PhD project investigates how structured reasoning scaffolds can improve the reliability and interpretability of LLM-generated judicial decisions. We develop FIRAC-segmented datasets of Indian judgments and explore their use in structured prompting and multi-agent pipelines for semi-transparent judgment generation. Early experiments show that FIRAC-based inputs improve predictive consistency, and we introduce metrics - covering lexical similarity, semantic alignment, and citation accuracy - to assess reasoning quality. Ongoing work focuses on establishing principled evaluation frameworks for legal reasoning and on integrating retrieval and knowledge-graph grounding to enhance legal soundness. We seek feedback on reasoning evaluation, transparency methods, and generalisable approaches for structured judicial decision generation within IR and AI for Law.