Modeling the Diachronic Evolution of Legal Norms: An LRMoo-Based, Component-Level, Event-Centric Approach to Legal Knowledge Graphs

Technical Summary: Modeling the Diachronic Evolution of Legal Norms

Problem Statement
The paper addresses a critical gap in the automated processing of legal norms: the inability of current systems to perform deterministic, point-in-time reconstruction of legal texts. While foundational frameworks like IFLA LRMoo and standards like Akoma Ntoso exist, they lack a formal pattern for granular, component-level versioning. This deficiency is particularly acute for Generative AI and Large Language Models (LLMs), which operate on probabilistic bases and lack the deterministic mechanisms required for high-stakes legal domains where correctness is absolute. Existing models often treat amendments as document-embedded annotations or separate consolidated snapshots, failing to explicitly model the causal lineage of individual components or the legislative events that drive change. Consequently, retrieving the exact text of a law as it existed on a specific past date remains unreliable.

Methodology
The authors propose a structured, temporal modeling pattern grounded in the IFLA Library Reference Model Ontology (LRMoo), extended to support component-level granularity and event-centric provenance. The methodology is built on three core pillars:

1. Two-Tier Versioning (TV/LV Distinction):

   • Temporal Version (TV): Modeled as a distinct F1 Work, representing the language-agnostic, semantic content of a norm at a specific point in time. Each legislative amendment creates a new TV, forming a chain of derivative Works.
   • Language Version (LV): Modeled as an F2 Expression, representing the concrete monolingual textual realization of a TV. This separation allows the model to handle multilingual jurisdictions (multiple LVs per TV) while maintaining a single semantic anchor for the legal intent.

2. Event-Centric Change Modeling:

   • Legislative amendments are not modeled as simple binary relations (e.g., "amends") but as first-class events.
   • A textual amendment is formalized as a complex event jointly typed as F27 Work Creation (creating the new TV) and E64 End of Existence (terminating the previous TV).
   • The event explicitly links the amending provision (source) to the affected norm (target) via properties such as P16 used specific object, P93 took out of existence, and R16 created.
   • Validity intervals are not static attributes but are derived dynamically from the timestamps of these bounding events (creation and termination), supporting both transaction time and valid time (including vacatio legis and retroactive effects).

3. Recursive Component-Level Granularity:

   • The model applies the same LRMoo pattern recursively to internal legal structures (articles, paragraphs, etc.).
   • Component Concepts (CC) and Component Temporal Versions (CTV) are modeled as F1 Works, linked via R67 has part and R2 is derivative of.
   • Amendments are decomposed into macro-events (the enactment of the amending norm) and micro-events (specific instructions affecting individual components). Each micro-event traces the precise causal link between a specific instruction in an amending act and the resulting change in a specific component's version.

Key Contributions

• Deterministic Reconstruction Architecture: The paper provides a formal structure that enables the exact reconstruction of any part of a legal text as it existed on any given date, offering a verifiable semantic backbone for Legal Knowledge Graphs (LKGs).
• Granular Component Versioning: By establishing parallel hierarchies for abstract components and their versions, the model moves beyond document-level versioning to track the diachronic history of individual legal provisions.
• Reified Event Provenance: The approach elevates legislative events to first-class entities, allowing for precise tracing of the provenance of every legal change, distinguishing between insertion, repeal, and wording amendments through specific ontological typing.
• Separation of Semantic Content and Linguistic Form: The clear distinction between language-agnostic Temporal Versions and monolingual Language Versions provides a robust mechanism for managing multilingual legal data without conflating semantic evolution with translation.

Results and Case Study
The authors demonstrate the model's efficacy using the Brazilian Federal Constitution of 1988 as a case study. The model successfully traces the Constitution's lifecycle through over one hundred amendments.

• It maps the initial 1988 promulgation to a specific TV and LV.
• It models the 1992 Constitutional Amendment (EC1) as a complex event that terminates the initial TV and creates a new TV, explicitly linking the amending instrument to the resulting text.
• It extends this logic to the component level, tracing the specific addition of the word "housing" (moradia) to Article 6 via Constitutional Amendment No. 26 (2000). The model captures this as a micro-event that terminates the previous Component Temporal Version (CTV0) and creates a new one (CTV1), preserving a complete, auditable lineage.

Significance and Claims
The paper positions this model as a necessary prerequisite for trustworthy Legal AI. It argues that while probabilistic models are insufficient for legal domains requiring absolute precision, a robust descriptive model based on LRMoo provides the "verifiable ground truth" essential for reliable Legal Knowledge Graphs. The significance lies in shifting the paradigm from document-centric metadata to a deterministic, event-sourced ontology that treats legal norms as evolving chains of intellectual works. This approach enables the construction of pipelines capable of answering "what was the law on date X?" with mathematical certainty, thereby supporting transparency, efficient legal research, and the development of reliable computational law applications. The authors conclude that this model serves as a foundational layer upon which advanced reasoning and benchmarks for legal AI can be built.