Abstract
<title>Abstract</title> <p>While legal judgment prediction systems demonstrate increasingly high accuracy, their practical deployment often faces severe normative challenges regarding automated adjudication and remains largely confined to binary classification tasks. This study bridges the gap between technical natural language processing capabilities and practical legal workflows by developing a retrieval-augmented language model tailored for pre-litigation alternative dispute resolution in high-volume mass disputes. Focusing on German road traffic liability, the proposed system predicts continuous, proportional liability distributions based on abbreviated factual summaries drawn from a leading legal commentary. An empirical evaluation examines diverse open-source large language models across different retrieval strategies alongside comprehensive ablation studies on the volume and ordering of in-context demonstrations. Experimental results reveal that integrating taxonomic legal metadata into the retrieval pipeline significantly outperforms purely semantic approaches. Furthermore, providing relevance-ordered demonstrations effectively mitigates context oversaturation, a vulnerability explicitly exposed when reference cases are randomly shuffled. The model achieves predictive error margins that align closely with the underlying operational risk doctrine of German traffic law, effectively capturing fault-based allocations. By transforming an authoritative legal text into a computable coordination device, this research demonstrates how algorithmic prediction can rationalize litigant expectations, alleviate information asymmetries, and facilitate out-of-court settlements without compromising substantive justice in civil law jurisdictions.</p>