SAMVAD: Simulated Agent Verdict System
- SAMVAD is a multi-agent system designed to simulate judicial deliberation in India using role-specialized agents and retrieval-augmented generation.
- The system integrates a domain-specific legal knowledge base with structured case files to produce transparent, legally grounded outcomes with verifiable citations.
- SAMVAD employs an iterative consensus protocol among judge, prosecution, defense, and adjudicator agents to model realistic legal reasoning and decision convergence.
SAMVAD, expanded as Simulated Agent-based Multi-agent Verdict Adjudication, is a Multi-Agent System (MAS) designed to simulate judicial deliberation within the Indian legal context by combining role-specialized LLM agents with Retrieval-Augmented Generation (RAG) grounded in a domain-specific legal knowledge base (Devadiga et al., 4 Sep 2025). The system models a Judge, a Prosecution Counsel, a Defense Counsel, and multiple Adjudicators simulating a judicial bench, and uses structured deliberation rounds, source-cited legal retrieval, and an orchestrated consensus protocol to study legal reasoning, deliberation quality, and outcome consistency. Its stated emphasis is on transparency, legal fidelity, and explainability, particularly through verifiable citations to Indian legal materials and logged deliberation traces (Devadiga et al., 4 Sep 2025).
1. System definition and architectural composition
SAMVAD is presented as an innovative Multi-Agent System for simulating the deliberation process in the framework of the Indian justice system (Devadiga et al., 4 Sep 2025). Its architecture comprises four primary elements: an Orchestrator as the central controller of simulation flow and communication, a set of role-specialized Agents, a Legal Knowledge Base (Vector store) for domain-specific retrieval, and an LLM Interface that connects the agents to selected LLMs (Devadiga et al., 4 Sep 2025).
The agent set is explicitly differentiated by judicial function. The system includes a Judge, a Prosecution Counsel, a Defense Counsel, and multiple Adjudicators that simulate a judicial bench (Devadiga et al., 4 Sep 2025). This role specialization is central to the design: the Judge produces legal instructions, counsel agents generate adversarial arguments, and Adjudicators deliberate iteratively before converging on a verdict. The architecture therefore separates procedural functions that, in human judicial settings, would ordinarily be distributed across institutional actors.
Case input is standardized through a Case File (JSON) format containing case_id, factual summary, charges, simplified laws, prosecution and defense evidence, and legal keywords (Devadiga et al., 4 Sep 2025). This structured representation constrains the simulation to a consistent interface across cases and supports reproducible orchestration. A plausible implication is that the system is intended less as a free-form legal chatbot than as a configurable experimental platform for controlled studies of judicial-panel dynamics.
2. Agent roles and deliberative behavior
The Judge Agent is defined as an impartial source of legal instructions for the Adjudicators (Devadiga et al., 4 Sep 2025). It queries the legal knowledge base for relevant legal principles and standards of proof, then outputs structured, citable instructions synthesizing legal context and case specifics, described as analogous to a “charge to the jury” (Devadiga et al., 4 Sep 2025). Prompt engineering is used to enforce neutrality and clarity.
The Prosecution Counsel and Defense Counsel agents are symmetrical in structure but adversarial in objective. Each uses RAG to retrieve statutes, definitions, and precedents supporting its side, then produces persuasive, law-grounded arguments with authoritative citations (Devadiga et al., 4 Sep 2025). Their outputs are therefore not merely rhetorical; they are intended to be explicitly grounded in retrieved legal text chunks from the domain corpus.
The Adjudicator Agents constitute the deliberative core of SAMVAD. In each round, every Adjudicator states a position—Guilty, Not Guilty, or Undecided—together with a reasoned justification (Devadiga et al., 4 Sep 2025). In later rounds, each Adjudicator reads peer justifications and may evolve its stance. This introduces explicit peer-to-peer influence into the simulation and is described as mimicking realistic, interactive group deliberation (Devadiga et al., 4 Sep 2025).
A common misconception would be to treat the Adjudicators as independent one-shot classifiers. The design instead defines them as participants in an iterative synchronous protocol in which visibility of peer reasoning is an intended mechanism of opinion revision. The system therefore simulates not only verdict formation but also the dynamics of convergence.
3. Retrieval-Augmented Generation and legal grounding
A primary contribution of SAMVAD is its integration of Retrieval-Augmented Generation grounded in a domain-specific knowledge base of landmark Indian legal documents, including the Constitution of India, the Indian Penal Code (IPC), and the Code of Criminal Procedure (CrPC) (Devadiga et al., 4 Sep 2025). The knowledge-base construction pipeline extracts PDF text, chunks it into segments while retaining section metadata, embeds those chunks with a sentence-embedding model, and stores the resulting dense vectors in ChromaDB for semantic retrieval (Devadiga et al., 4 Sep 2025).
The RAG workflow begins with an agent’s information need, such as a query about the standard of proof for a statutory provision, which is embedded and used for top-k similarity search in the vector store (Devadiga et al., 4 Sep 2025). Retrieved chunks are then supplied to the LLM alongside the agent prompt, after which the model generates output that references legal sources with explicit provenance, including section and document metadata (Devadiga et al., 4 Sep 2025). The paper states that this mechanism allows Judge and Counsel agents to generate legally sound instructions and arguments, complete with source citations, thereby enhancing both fidelity and transparency (Devadiga et al., 4 Sep 2025).
This legal grounding is treated as a safeguard against hallucinated legal reasoning. The evaluation criteria explicitly define high legal grounding as arguments that cite relevant law or precedent rather than hallucinated material (Devadiga et al., 4 Sep 2025). A plausible implication is that SAMVAD’s explainability claims depend less on the intrinsic reliability of the underlying LLMs than on the externalization of supporting legal sources through retrieval and citation.
4. Simulation workflow and consensus protocol
The simulation workflow proceeds in five phases (Devadiga et al., 4 Sep 2025). First, during Initialization, the system loads the vector-store knowledge base, the case JSON, and instantiates all agents. Second, in Trial Preparation, the Judge uses RAG to draft impartial legal instructions while both Counsel agents use RAG to draft their arguments, each citing a legal basis. Third, during Jury Deliberation (Iterative Rounds), the Adjudicators receive case facts, judicial instructions, and counsel arguments, and formulate positions with justifications in round 1; in later rounds, they may revise their positions after examining peer statements. Fourth, in Consensus Check, the Orchestrator evaluates whether a predefined agreement threshold has been reached. Fifth, in Conclusion and Reporting, the system compiles the final verdict, deliberation traces, metrics, citations, and an explainability report (Devadiga et al., 4 Sep 2025).
The communication protocol is synchronous. Adjudicators communicate their leanings and justifications in parallel, and all statements in a round become visible before the next round starts (Devadiga et al., 4 Sep 2025). This protocol is intended to promote opinion evolution through exposure to peer reasoning.
Consensus is operationalized quantitatively. The paper specifies that if
then consensus is declared; otherwise, additional rounds occur until either consensus is reached or the round limit is exceeded, in which case a hung panel may be declared (Devadiga et al., 4 Sep 2025). The accompanying pseudocode expresses the same rule as a mode-frequency threshold check over adjudicator leanings (Devadiga et al., 4 Sep 2025).
This threshold-based rule is important because it converts deliberation into a measurable stopping condition. It also makes bench size, round limits, and model choice configurable, which the paper identifies as part of the system’s broader configurability (Devadiga et al., 4 Sep 2025).
5. Evaluation framework and reported simulation outcomes
SAMVAD is accompanied by a comprehensive evaluation plan intended to assess performance, deliberation quality, and outcome consistency (Devadiga et al., 4 Sep 2025). The reported metrics are: Performance, defined as agent LLM response times with mean, min, and max; Participation Rate, defined as the fraction of Adjudicators contributing in each round; Argument Grounding Score, defined as the fraction of key legal or case keywords present in agent statements; Meaningful Statements, defined as the proportion of Adjudicator responses with substantive justification; Agreement Ratio, defined as the final consensus fraction; and Verdict Consistency, defined as reproducibility of verdicts across simulation runs, measured by majority verdict rate (Devadiga et al., 4 Sep 2025).
The reported case-study results cover five cases, each with 5 Adjudicators and RAG enabled for both Judge and Counsel agents (Devadiga et al., 4 Sep 2025). The Final Verdicts are reported as NG, NG, G, NG, G across Cases 1 through 5. Deliberation Rounds are 1, 1, 2, 2, 1, respectively. Final Agreement Ratio is 1.00, 1.00, 0.90, 1.00, 0.95. Participation Rate is 1.00 in all cases. Avg. Argument Grounding is 0.32, 0.30, 0.45, 0.42, 0.38 (Devadiga et al., 4 Sep 2025).
These results are summarized below.
| Metric | Reported values |
|---|---|
| No. of Adjudicators | 5 for Cases 1–5 |
| Final Verdict | NG, NG, G, NG, G |
| Deliberation Rounds | 1, 1, 2, 2, 1 |
| Final Agreement Ratio | 1.00, 1.00, 0.90, 1.00, 0.95 |
| Participation Rate | 1.00 in all cases |
| Avg. Argument Grounding | 0.32, 0.30, 0.45, 0.42, 0.38 |
The paper characterizes these results as showing high participation and strong consensus, with some cases resolving in one round and others requiring two rounds, reflecting case complexity (Devadiga et al., 4 Sep 2025). Because these are simulation outcomes rather than courtroom observations, their significance lies in internal system behavior: they indicate that the orchestration protocol can sustain full participation and high agreement under the tested configuration.
6. Ablation, interpretive significance, and scope
The ablation study compares two model families—Qwen-2.5 7B and LLaMA-3 8B—with and without RAG (Devadiga et al., 4 Sep 2025). For Qwen-2.5 7B, enabling RAG yields Agreement 0.98, Grounding 0.42, and Consistency Very High, whereas disabling RAG yields Agreement 0.90, Grounding 0.21, and Consistency Medium. For LLaMA-3 8B, enabling RAG yields Agreement 0.95, Grounding 0.37, and Consistency High, whereas disabling RAG yields Agreement 0.87, Grounding 0.18, and Consistency Low (Devadiga et al., 4 Sep 2025). The paper explicitly states that grounding scores nearly double with RAG and that gains persist across different LLM families, indicating a model-agnostic RAG benefit (Devadiga et al., 4 Sep 2025).
The significance attributed to SAMVAD rests on three linked properties. First, legal fidelity is pursued by grounding arguments and instructions in authoritative Indian legal texts (Devadiga et al., 4 Sep 2025). Second, transparency and explainability are pursued through citation tracing and full logging of deliberation traces for external review (Devadiga et al., 4 Sep 2025). Third, configurability is emphasized through adjustable bench size, rounds, LLM choice, and modular case injection (Devadiga et al., 4 Sep 2025). Together these design choices frame SAMVAD as a research platform for exploring legal reasoning and group decision-making dynamics rather than as an autonomous judicial authority.
A likely misconception is that simulation fidelity and legal correctness are identical. The paper does not claim that SAMVAD replaces judicial institutions or that its verdicts constitute legal judgments; rather, it presents a configurable and explainable MAS platform for studying how deliberation might unfold under structured legal grounding (Devadiga et al., 4 Sep 2025). This suggests that the system’s primary research value lies in controlled experimentation on deliberative behavior, reproducibility, and grounding, especially where empirical study of real judicial panels is constrained by ethical and practical barriers (Devadiga et al., 4 Sep 2025).
