Building Trustworthy Legal AI: A Framework for AI Act Compliance and Traceability

TLDR: A research paper introduces a “RAG-Forense” framework for AI systems in the legal sector to ensure verifiable compliance with the EU AI Act. It features a forensic architecture with cryptographic traceability, temporal validity, policy-as-code guardrails, and secure logging. Experimental results show this framework significantly outperforms standard AI models in accuracy, safety, and auditability, providing a blueprint for responsible legal AI development.

The rapid advancement of Artificial Intelligence (AI) is transforming various sectors, and the legal field is no exception. From drafting assistants to systems aiding judicial authorities, AI offers immense potential. However, this also introduces significant risks, particularly concerning accuracy, reliability, and accountability. The European Union’s AI Act (Regulation (EU) 2024/1689) aims to address these challenges by establishing a comprehensive legal framework for AI, especially for high-risk applications like those in the legal sector.

A recent research paper, titled “Gobernanza y trazabilidad ‘a prueba de AI Act’ para casos de uso legales: un marco técnico-jurídico, métricas forenses y evidencias auditables” (AI Act-Ready Governance and Traceability for Legal Use Cases: A Techno-Legal Framework, Forensic Metrics, and Auditable Evidence), by Alex Dantart of LittleJohn, proposes a groundbreaking framework to ensure AI systems in the legal domain are not just functional, but also verifiably compliant with the stringent requirements of the AI Act. This work is crucial as the AI Act’s obligations for General Purpose AI (GP AI) models come into effect in August 2025, followed by most high-risk systems in August 2026.

The Core Challenge: Beyond Declarative Policies

The central challenge highlighted by the paper is that demonstrable compliance with the AI Act demands more than just written policies. It requires end-to-end traceability of AI decisions, temporal accuracy of legal norms, robust version control, continuous governance, and metrics that truly reflect the legal cost of errors. A major concern in AI, especially with large language models (LLMs), is ‘hallucinations’ – the generation of factually incorrect or misleading content. In law, where fidelity to sources is paramount, hallucinations pose a systemic risk.

Introducing RAG-Forense: A Forensic Architecture for Legal AI

The paper introduces a comprehensive framework that integrates compliance “by design” into the AI system’s lifecycle. This framework culminates in a technical architecture called “RAG-Forense,” which stands for Retrieval-Augmented Generation (RAG) with forensic capabilities. Unlike generative AI, which acts as a ‘creative oracle,’ RAG-Forense operates as an ‘expert archivist,’ focusing on retrieving, structuring, and presenting verified knowledge.

The RAG-Forense architecture is built upon six key principles:

• End-to-End Forensic Traceability: Every output can be traced back to its original source, including the exact versions of documents used. This directly addresses the AI Act’s requirement for record-keeping (Article 12).
• Regulatory Temporality: The system can reason about the state of the law at a specific point in time, preventing the use of outdated or repealed regulations (Articles 10 and 15).
• Anchored Citation: All legal statements are directly supported by literal, verifiable passages from the source corpus, acting as a primary defense against hallucinations (Article 13).
• Calibrated Abstention: The system is designed to abstain from answering when evidence is insufficient, the question is outside its scope, or improper legal advice is requested. This threshold is calibrated based on legal risk (Articles 9 and 14).
• Defense in Depth: Security and robustness are implemented in multiple layers to protect against prompt injection attacks and data leakage (Article 15).
• Continuous Governance: The framework generates artifacts for ongoing audits, risk monitoring, and lifecycle management, aligning with international standards like ISO/IEC 42001.

How RAG-Forense Works

The technical architecture involves a four-stage process:

1. Secure Ingestion and Versioning: Legal documents are processed to ensure integrity and temporality, assigned cryptographic hashes, and stored in a Write-Once, Read-Many (WORM) repository to prevent unauthorized changes.
2. Indexing with Temporal Partitioning: Documents are indexed and vectorized, but the index is logically partitioned by date ranges. This means a query with a specific date context will only search documents valid at that time.
3. Orchestration with Policy-as-Code (PaaC): An orchestrator component applies explicit rules before generating a response. These rules verify if enough information is available, instruct the LLM to cite sources for every statement, and filter out prompt injection attempts or requests for improper legal advice.
4. Secure Logging WORM: Every interaction, including the user’s prompt, temporal context, document hashes, LLM response, and policy evaluations, is packaged, cryptographically signed, and written to an immutable log. This creates auditable evidence as required by the AI Act.

Experimental Validation and Key Findings

To validate the framework, the researchers compared three systems: an LLM-Only model, a standard RAG-Base system, and the proposed RAG-Forense. They used a synthetic corpus based on different versions of the AI Act itself, allowing for precise control over temporal accuracy.

The results were conclusive: RAG-Forense significantly outperformed both baseline systems across all compliance metrics. For instance, it achieved a 95% temporal validity (TV@date) compared to 49% for LLM-Only and 71% for RAG-Base. Its anchored citation precision (ACP) was 92%, and its unsafe advice rate plummeted to 0.7%. The system’s higher abstention rate (18.4%) was interpreted not as a weakness, but as a crucial safety mechanism, demonstrating its ability to defer ambiguous or high-risk cases to human professionals, aligning with the AI Act’s human oversight requirements.

The paper emphasizes that compliance with the AI Act is not an emergent property but a feature that must be deliberately designed. Standard RAG systems, while an improvement over pure LLMs, still fall short in critical areas like temporal validity and security, suggesting that many current legal tech implementations might not meet the AI Act’s rigorous demands.

Also Read:

• Building Secure AI Agents: Understanding the Trust, Risk, and Liability Framework
• Evaluating AI Performance in Finance: A New Framework for Mitigating Metric Risks

Looking Ahead: Operational Challenges and Future Directions

While the sandbox results are promising, the paper also discusses the challenges of transitioning RAG-Forense to real-world production. These include the continuous and verified curation of legal corpora, managing legal uncertainty and evolving doctrine, and optimizing scalability for forensic traceability. The authors propose solutions like automated source monitors, metadata extraction engines, and human-in-the-loop verification for corpus quality, as well as advanced Policy-as-Code rules for conflict management and tiered storage for forensic logs.

The research also highlights the critical role of the human factor, acknowledging that even with a robust AI, the risk of ‘user hallucination’ (uncritical trust in AI output) persists. The framework aims to empower legal professionals as critical supervisors, not replace them.

This work represents a significant step towards building genuinely reliable and accountable AI systems for the legal sector, providing a practical blueprint for compliance with the EU AI Act. The open-source release of rag-forense further contributes to this goal, offering a tangible tool for developing responsible legal tech. For more details, you can read the full research paper.