Legal judgments are among the densest and most consequential bodies of text that automated systems are asked to summarize, and the people who rely on those summaries – judges, lawyers, and researchers – cannot afford small errors in what gets kept or dropped. Two recent lines of work highlight both the promise and the limits of current approaches. Reference-free evaluation frameworks such as LEAD show that scoring legal summaries along relevance, coherence, consistency, and fluency using domain-adapted language models correlates far better with expert judgment than older overlap-based metrics. Separately, lightweight abstractive systems built on domain-tuned sequenceto-sequence models, such as legal-pegasus, generate fluent summaries quickly, though they sometimes rely on inserting fixed template sentences when certain keywords are missing from the output – a shortcut that can introduce content the source document never actually supports. Neither line of work closes the loop between generating a summary and verifying that every sentence in it is actually grounded in the judgment. This paper introduces FAITH-LEGAL, a retrieval-grounded, citation-aware summarization pipeline that traces each generated sentence back to specific evidence in the source judgment, checks that trace with a natural language inference model sensitive to both contradiction and lack of support, and regenerates or removes any sentence that fails the check – rather than patching it with an unverified template. We also extend the LEAD framework into LEAD+, adding a Citation Grounding Score, and introduce the Unsupported Content Rate (UCR) as a direct, sentence-level measure of hallucination. We detail the full architecture, the mathematical formulation behind each scoring component, and the experimental protocol used to evaluate the system against extractive and abstractive baselines drawn from prior work, including an ablation study isolating the contribution of retrieval, citation alignment, and the correction loop.

legal,Automatic Summarization,NLP,Machine Translation; Evaluation Methods; Faithfulness; Intelligibility,bert,retrieval-augmented generation,citation-aware,InLegalBERT,natural language inference.