为解决法律研究中法条历史沿革追踪困难及大型语言模型（LLM）在时效性处理上的缺陷，开发者推出了一种基于Git逻辑的美国法典版本控制系统。1 该系统的核心痛点在于律师和AI模型难以便捷获取法条的完整历史信息，包括制定者、生效时间及过往状态。1 针对这一挑战，项目团队开发了可视化工具，允许用户查看特定日期的生效文本并追溯变更来源，旨在通过提供可检查的历史记录来增强AI辅助法律研究的准确性。1
在数据验证方面，该项目对比了2017年以来的发布点，结果显示365项修改中有361项可精确复现。1 关于法条生效日期的分类，统计数据显示有261项默认生效，72项依据生效日期注释，另有29项依据修正案内部日期。1 以《美国法典》第11编第1182(1)节为例，“小企业债务人”定义中的债务限额曾从750万美元临时上调至2022年，随后恢复又过期，并在2025年调整后为342.4万美元，这类复杂变动凸显了精确版本控制的必要性。1
研究进一步揭示了现有LLM在法律应用中的局限性。慕尼黑工业大学的研究发现，LLM常应用已过时的规则或在旧法适用时偏好新法；单独回答截止后问题的正确率仅为24%，而结合检索后升至87%。1 尽管尝试让模型应用修正案生成旧文本时相似度可达93-99%，但单条修正案匹配率仅50%，存在显著的法律错误风险。1 目前该系统仍存在局限，尚未完全处理后续法律追溯覆盖早期日期的情况，例如2022年6月法案对2022年3月前案件的追溯力显示不足。1
Researchers have developed a version control system modeled after Git to track the historical evolution of the United States Code, aiming to address significant challenges in legal research and improve the accuracy of Large Language Models (LLMs). 1 The core issue driving this project is that both lawyers and AI models struggle to access the complete history of statutes, including who authored them, when they took effect, and their previous states. 1 This lack of accessible temporal data causes LLMs to frequently confuse old and new laws, as demonstrated by findings from the Technical University of Munich which showed that models often apply outdated rules or prefer newer laws even when older ones are applicable. 1
To validate the system’s reliability, developers tested it against 365 legislative changes occurring since 2017, successfully reproducing 361 of these modifications at their respective release points. 1 These changes were categorized by effective date logic: 261 defaulted to immediate effectiveness, 72 relied on explicit effective date annotations, and 29 depended on dates specified within the amendment itself. 1 A specific example illustrates the complexity of tracking these changes: the definition of "small business debtor" under Title 11 USC § 1182(1) saw the debt limit temporarily raised to $7.5 million until 2022, reverted, expired, and was subsequently adjusted to $3,424,000 in 2025. 1
The tool also seeks to mitigate risks associated with automated merging of legal texts, where models might incorrectly apply amendments despite high textual similarity scores ranging from 93% to 99%. 1 Research indicates that while overall text similarity is high, the matching rate for individual amendments can drop to 50%, introducing substantial potential for legal errors if not carefully constrained. 1 When isolated questions about post-cutoff dates are asked, LLM accuracy drops to just 24%, but integrating retrieval mechanisms boosts this figure to 87%, highlighting the necessity of providing checkable historical records. 1 Visualization tools allow users to view the text effective on specific dates and identify the source of changes, though the system currently has limitations in handling retroactive coverage, such as displaying insufficient information for cases affected by June 2022 legislation applied to pre-March 2022 events. 1
评论
还没有评论，欢迎留下第一条。