Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reliability by design: quantifying and eliminating fabrication risk in LLMs. From generative to consultative AI: a comparative analysis in the legal domain and lessons for high-stakes knowledge bases

Alex Dantart

TL;DR

This study distinguishes generative and consultative AI paradigms in high-stakes legal drafting and introduces two reliability metrics, FCR and FFR, to quantify factual integrity. Using the JURIDICO-FCR dataset (75 tasks across Spanish law) and 2,700 runs with 12 LLMs, it demonstrates that pure generative AI yields substantial hallucinations, while canonical RAG substantially reduces errors but leaves non-negligible misgrounding. Advanced RAG—featuring domain-specific embeddings, multi-step retrieval, re-ranking, and a verification/self-correction loop—drastically minimizes fabrication to near-negligible levels, establishing a practical path to reliable legal AI. The findings advocate for a shift toward rigor-focused, traceable consultative architectures and outline an evaluation framework applicable to other high-risk domains, promoting a culture of informed skepticism and human-centered verification.

Abstract

This paper examines how to make large language models reliable for high-stakes legal work by reducing hallucinations. It distinguishes three AI paradigms: (1) standalone generative models ("creative oracle"), (2) basic retrieval-augmented systems ("expert archivist"), and (3) an advanced, end-to-end optimized RAG system ("rigorous archivist"). The authors introduce two reliability metrics -False Citation Rate (FCR) and Fabricated Fact Rate (FFR)- and evaluate 2,700 judicial-style answers from 12 LLMs across 75 legal tasks using expert, double-blind review. Results show that standalone models are unsuitable for professional use (FCR above 30%), while basic RAG greatly reduces errors but still leaves notable misgrounding. Advanced RAG, using techniques such as embedding fine-tuning, re-ranking, and self-correction, reduces fabrication to negligible levels (below 0.2%). The study concludes that trustworthy legal AI requires rigor-focused, retrieval-based architectures emphasizing verification and traceability, and provides an evaluation framework applicable to other high-risk domains.

Reliability by design: quantifying and eliminating fabrication risk in LLMs. From generative to consultative AI: a comparative analysis in the legal domain and lessons for high-stakes knowledge bases

TL;DR

This study distinguishes generative and consultative AI paradigms in high-stakes legal drafting and introduces two reliability metrics, FCR and FFR, to quantify factual integrity. Using the JURIDICO-FCR dataset (75 tasks across Spanish law) and 2,700 runs with 12 LLMs, it demonstrates that pure generative AI yields substantial hallucinations, while canonical RAG substantially reduces errors but leaves non-negligible misgrounding. Advanced RAG—featuring domain-specific embeddings, multi-step retrieval, re-ranking, and a verification/self-correction loop—drastically minimizes fabrication to near-negligible levels, establishing a practical path to reliable legal AI. The findings advocate for a shift toward rigor-focused, traceable consultative architectures and outline an evaluation framework applicable to other high-risk domains, promoting a culture of informed skepticism and human-centered verification.

Abstract

This paper examines how to make large language models reliable for high-stakes legal work by reducing hallucinations. It distinguishes three AI paradigms: (1) standalone generative models ("creative oracle"), (2) basic retrieval-augmented systems ("expert archivist"), and (3) an advanced, end-to-end optimized RAG system ("rigorous archivist"). The authors introduce two reliability metrics -False Citation Rate (FCR) and Fabricated Fact Rate (FFR)- and evaluate 2,700 judicial-style answers from 12 LLMs across 75 legal tasks using expert, double-blind review. Results show that standalone models are unsuitable for professional use (FCR above 30%), while basic RAG greatly reduces errors but still leaves notable misgrounding. Advanced RAG, using techniques such as embedding fine-tuning, re-ranking, and self-correction, reduces fabrication to negligible levels (below 0.2%). The study concludes that trustworthy legal AI requires rigor-focused, retrieval-based architectures emphasizing verification and traceability, and provides an evaluation framework applicable to other high-risk domains.
Paper Structure (35 sections, 2 equations, 2 figures, 5 tables)

This paper contains 35 sections, 2 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Extrapolation to Other Domains or Sectors
  3. Related work
  4. The nature of hallucinations: system failure or design feature
  5. The human factor: cognitive biases and the acceptance of errors
  6. Fact-Checking and the challenge of legal Ground Truth
  7. Evaluating AI in the legal domain: from reasoning benchmarks to reliability in drafting
  8. RAG as an implementation of the consultative AI paradigm
  9. Canonical (or Basic) RAG
  10. Advanced RAG (fusion architecture and deep verification):
  11. A framework for evaluating factual integrity in legal texts
  12. Fundamental definitions
  13. Quantitative metrics
  14. Qualitative taxonomy of errors
  15. Experimental setup
  16. ...and 20 more sections

Figures (2)

  • Figure 1: Comparison of average FCR with simulated empirical variability. Note: To properly visualize the magnitude differences between paradigms, the numerical values above the bars represent the natural logarithm of the rate (ln(FCR)), where negative values indicate rates close to zero. The logarithmic axis continues to show a drastic reduction in error with RAG architectures, while the nonlinear data reflect a more realistic test scenario.
  • Figure 2: Comparison of average FFR. The values represent the natural logarithm of the percentage ln(%) to highlight the scale of the reduction. The non-linearity in model performance is evident, but the superiority of the Advanced RAG architecture in minimizing fact fabrication remains the dominant conclusion.