JUDICATA: An Agentic Neuro-Symbolic Architecture for Legal Reasoning with Adversarial Defense and Procedurally Contestable Argumentation

<title>Abstract</title> <p>Retrieval-Augmented Generation (RAG) systems designed for the legal domainexhibit hallucination rates ranging from 17% to 33%, even when grounded incurated databases such as Westlaw and LexisNexis. General-purpose languagemodels reach failure rates of 69%–88% on legal queries. Such findings indi-cate that document retrieval alone is insufficient to ensure the reliability ofautomated legal reasoning. This work introduces JUDICATA (Judicial Defenseand Intelligent Causal Architecture for Trustworthy Argumentation), an agenticframework that integrates five core components: (1) HyPA-RAG, an adaptivetri-modal retrieval mechanism with query complexity-aware calibrated weights;(2) LSIM, a reasoning layer based on Fact–Rule–Conclusion chains that com-plement probabilistic generation with explicit logical structure; (3) GuardianDual-Pass, a two-stage defense combining syntactic analysis and semantic simi-larity filtering (τ = 0.84); (4) LOPSIDED, a pseudonymization mechanism thatpreserves narrative coherence; and (5) structured output via the Toulmin Modelwith 6 components, enabling procedural contestability. Experimental evaluationcomprises three LLMs (local Mistral 7B, Kimi K2 via Groq, Gemini 2.5 Flashvia Vertex AI) across three dimensions: security (Guardian: 755/800 attacksdetected, 94.4%, with 0 False Positives (FP) on n = 100 queries; P2P: 95.0%;Anonymizer: 97.0%), reasoning quality (Mistral 7B: Fact Extraction 205/210,Toulmin 68/70 with composite score 0.798; LSIM 67/70 with composite 0.605.Kimi K2: baseline n = 630, ablation n = 300), and robustness (edge cases97/101, temporal stability). An ablation study (n = 50 per configuration, sixconfigurations) identifies the Toulmin module as a critical component: its removalreduces the composite score by −0.174 (Kimi K2) and −0.185 (Mistral 7B),while hallucination risk increases to 0.97. Baseline comparisons (raw LLM vs.vanilla RAG vs. JUDICATA) demonstrate consistent performance improvementsacross both models. The primary contribution of this work does not lie in novelalgorithms, but in the architectural integration of established techniques suchthat the resulting system exhibits adversarial defense, structured reasoning, andprocedural contestability—properties not achieved by any individual compo-nent in isolation. The framework is model-agnostic and operates seamlessly withboth local and remote services, providing robust legal reasoning independent ofproprietary cloud APIs when deployed in sovereign environments.</p>