Explainable Predictive Jurisprudence for Anticipating Legal Doctrine Evolution in Dynamic Judicial Systems

Abstract

Legal systems evolve continuously in response to legislative reforms, emerging judicial interpretations, and shifting societal expectations, making it increasingly difficult to anticipate changes in legal precedent using conventional analytical methods. This study introduces an explainable artificial intelligence framework for predictive jurisprudence that captures the temporal evolution of legal reasoning by jointly modeling semantic, structural, and causal relationships within judicial decisions. The proposed framework integrates neural temporal graph networks to learn evolving citation dependencies, dynamic topic modeling to characterize changes in legal doctrines over time, and causal-inference techniques to distinguish genuine jurisprudential influence from spurious associations. To enhance transparency, the predictive process is complemented by GNNExplainer, enabling the identification of the legal principles, precedents, and citation patterns that most strongly influence model predictions. The framework is evaluated using the Free Law Project and LePaRD benchmark datasets and demonstrates superior performance over existing approaches in detecting causal judicial influences and accurately quantifying precedent evolution. Its practical applicability and interpretability are further validated through expert legal assessment and historical backtesting against documented jurisprudential shifts. The experimental findings demonstrate that integrating explainable machine learning with causal legal analytics provides reliable early indicators of doctrinal change, offering valuable decision-support capabilities in legal environments.

1. Introduction

The philosophy of legislation is dynamic and evolves over time through legislative amendments in response to different events and societal changes. Case law is a fundamental element of public law frameworks. It is composed of networks of influence with different layers in which previous decisions control upcoming rules. However, the majority of computational paradigms consider the legal codes as static information, which engenders ignoring temporal law reasoning links.

This recent fast progress in artificial intelligence paradigms has allowed for better processing of natural languages and graphs, which has enhanced the management of difficult legal cases. In this regard, transformer-based language models are currently utilized to analyze legal textual content. In the same context, graph neural networks [1] are increasingly automated to represent citation networks. However, these advances remain restricted in interpreting critical legal decisions, providing causal explanations, and considering temporal characteristics.

To address the challenge of tracking temporal developments, this research proposes a new framework for predictive jurisprudence. This framework considers the concept of precedent drift as the simultaneous evolution of legal systems through causal, semantic, and structural dimensions. In this way, it becomes possible to move beyond simply analyzing past legal changes to predicting future jurisprudential shifts by integrating causal reasoning, dynamic thematic modeling, and time-based graph learning.

Transparency is a relevant issue in legal AI. Closed models [2] in legal contexts need explanations that should be aligned with case law. To address this issue, our system involves explanations using the GNNExplainer [3] model to classify important cases.

The suggested model is assessed on large-scale legal datasets, such as Free Law Project and LePaRD. The model enables both a high predictive accuracy and a practical vision for legal practitioners and policymakers.

The contributions of this study are as follows:

-

A unified precedent drift predictive model combining causal, semantic, and structural legal changes.

-

Hybridizing this model with TGNN, DTM, and causal inference for forecasting.

-

Adding an explainable layer via GNNExplainer, for identifying and analyzing precedent drift.

-

Experimenting with the proposed model and comparing with standard algorithms on two datasets (Free Law Project and LePaRD).

The paper raises the following research questions:

-

How does one model and enumerate precedent drifts in evolving legal applications?

-

Is explainable AI able to enhance explainability in predictive legal models?

-

How does causal inference distinguish true precedent influences from correlations?

-

How does temporal graph learning, semantic and causal modeling optimize the prediction of legal cases?

2. Related Works

Despite the excessive progress achieved in the field of using artificial intelligence techniques in analyzing linguistic texts, legal reasoning, learning dynamic systems, and studying graphs, most of the current smart technologies are unable to simultaneously process different dimensions, such as causal, semantic, and structural dimensions, in one problem.

2.1. Summary of State of the Art

The study in [4] presents a solution for predicting legal citations based on temporal graph neural networks (TGNNs). Compared to standard GNNs, this system achieves better recognition of local and global structural patterns. Despite this, this system ignores semantic data extracted from legal texts and causal inference. Besides, it does not explain citations, which is important in applicative contexts requiring interpretability, such as supporting legal decision-making.

In another study [5], time-based neural networks with attention mechanisms were proposed to generate timelines of events. This study accurately determines temporal dependencies and sequence patterns, even for heterogeneous datasets. However, the study ignores domain-specific inferences such as legal semantics, case law structures, or causal relationships, which minimizes the performance of modeling jurisprudence.

The study in [6] introduces a temporal heterogeneous GNN relying on patent data. The used GNN network combines semantic and structural features to predict technology convergence. However, this system lacks both causal inference and explainability. Moreover, it is considered a technological advance rather than a legally interpretable justified framework.

The authors in [7] implement a hyperbolic temporal GNN to design a hierarchical relationship for dynamic knowledge graphs. The model can efficiently represent complex relationships due to the use of hyperbolic spaces, which allow for better resolution of knowledge graph completion tasks. However, this system does not consider semantic drift or causal influence and is not dedicated to legal data since it focuses on learning reasoning rather than predictive reasoning.

The framework in [8] resolves a legal case retrieval problem using a graph structure with a set of textual features. The model concludes that hybridizing textual similarity with citations enhances the retrieval efficiency.

However, the model is static and does not account for temporal evolution. It also lacks predictive capabilities and causal reasoning, focusing instead on similarity-based retrieval rather than forecasting legal changes.

Table 1. Investigation of recent state-of-the-art methods.

Paper	Problem	Methodology	Complexity	Dataset(s)	Results	Pros	Cons	Comments
(Zhang et al., 2025) [4]	Citation prediction	Heterogeneous TGNN + clustering	O(E·d)	Academic graphs	Improved accuracy vs. GNN	Captures global + local structure	No causal modeling	Strong temporal modeling but domain-limited
(Aldawsari et al., 2026) [5]	Event timeline construction	Temporal GNN + attention	O(E·d)	Multi-domain events	Accuracy of 94.3%	Builds high temporal reasoning. Models causal order of events.	No semantic topic modeling	Focuses on events, not legal reasoning
(Li et al., 2025) [6]	Technology evolution	Temporal heterogeneous GNN	O(E·d)	Patent datasets	SOTA prediction	Combines structure or semantics	No explainability or causality	Closest to our hybrid idea
(Li et al., 2024) [7]	Knowledge graph completion	Hyperbolic TGNN	O(E·d)	Temporal KG benchmarks	SOTA performance	Models complex relations	No legal domain, no causality	Strong representation learning
(Tang et al., 2024) [8]	Legal case retrieval	Graph + text (CaseGNN)	O(E·d)	COLIEE datasets	Outperforms baselines	Uses graph-based structure and text for legal reasoning	Static (no temporal modeling)	Legal domain but not predictive
(Yuan et al., 2025) [9]	Community detection	Temporal GNN + attention	O(E·d)	Temporal networks	Improved accuracy	Captures dynamics	Not domain-specific	Generic TGNN model
(Ahmed et al., 2025) [10]	GNN survey	Comprehensive review	N/A	Multiple	N/A	Broad coverage	No unified framework	Highlights open challenges
(Liang et al., 2025) [11]	Spatiotemporal forecasting with interpretability and causal understanding in dynamic graph environments.	Dynamic causal explanation-based diffusion variational GNN (DCE-DVGNN)	High computational complexity due to graph diffusion operations. Complexity grows with graph size, time steps, and latent-variable optimization.	Spatiotemporal traffic/sensor datasets (METR-LA and PEMS).	Improved forecasting accuracy over baseline ST-GNN methods while causal explanations demonstrated better robustness and explainability.	Combines prediction and explainability. Handles dynamic spatial-temporal dependencies. Introduces causal interpretability.	High training cost. Complex architecture. Explainability quality may depend on causal assumptions. Reduced scalability for very large dynamic graphs.	Highly relevant for research combining XAI and temporal GNNs. Strong contribution for interpretable forecasting systems and causal graph learning.
(Srivastav et al., 2026) [12]	Modernization of criminal justice processes using AI-driven forensic and legal analytics.	Combines machine learning, NLP, and legal analytics concepts.	Moderate complexity overall. Complexity depends on underlying AI modules used in legal analytics and forensic processing.	Legal records from criminal justice and forensic evidence repositories.	Improved case processing and forensic analysis accuracy in justice systems.	Discusses ethical and legal implications. Connects AI with justice reform.	Limited experimental validation. Bias and privacy concerns. Legal explainability challenges remain.	Useful as a high-level reference for AI in legal systems. More conceptual than technically deep compared with GNN/XAI papers.
(Wang et al., 2026) [13]	Influence maximization in evolving temporal networks.	Combines Continuous-Time Graph Neural Networks (CT-GNNs) with Deep Reinforcement Learning (DRL)	Very high complexity due to continuous-time dynamic graph processing, reinforcement learning optimization, and sequential decision-making.	Temporal social network dataset, and diffusion benchmark datasets.	Superior influence spread and adaptive decision-making compared to classical influence maximization methods.	Captures temporal evolution accurately Integrates RL with GNNs	Harder interpretability. Computationally expensive. Scalability concerns on massive graphs.	Relevant to evolving legal and social network analysis.
(Zhu et al., 2026) [14]	Supply-chain forecasting with causal-aware dependency modeling.	Hybrid Graph Attention Networks and LSTM architecture.	Moderate-to-high complexity because of hybrid deep architecture involving attention	Supply-chain datasets, logistics/industrial forecasting datasets.	Better forecasting accuracy and causal-awareness compared to traditional forecasting and standalone DL approaches.	Captures spatial and temporal dependencies. Attention improves interpretability. Practical industrial relevance.	Requires large training data. Model tuning complexity. Limited causal certainty despite “causal-aware” design.	Strong industrial application of hybrid GNN-LSTM. Useful reference for causal-aware temporal forecasting architectures.
(Feng et al., 2023) [15]	Legal charge prediction from judicial documents while preserving semantic relationships between criminal actions.	Suggested Criminal Action Graph (CAG) representing legal facts as semantic graphs combined with deep learning/NLP.	Moderate-to-high complexity due to graph construction and semantic extraction.	Chinese judicial judgment datasets and criminal case documents.	Better legal charge prediction accuracy compared with traditional text-based legal NLP methods.	Captures semantic legal relations. Better contextual understanding. Shows the relevance of temporal structure in law.	Limited transferability across legal systems.	Useful reference for AI in legal analytics and graph-based legal reasoning.
(Zhang et al., 2026) [16]	Estimation of temporal centrality in evolving dynamic networks.	Temporal Graph Neural Networks (TGNNs) to estimate Temporal Katz Centrality efficiently in dynamic graphs.	High complexity due to temporal graph processing and recursive centrality estimation.	Dynamic graph benchmarks and temporal network datasets.	Accurate approximation of temporal Katz centrality with better scalability than classical computation methods.	Handles temporal evolution. Suitable for large-scale dynamic networks.	Computationally intensive training. Limited interpretability. Performance sensitive to temporal graph quality.	Relevant for temporal network analysis and evolving graph intelligence systems.
Proposed Model	Predictive jurisprudence	TGNN + DTM + causal + XAI	O(E·d)	Free Law Project, LePaRD	SOTA across tasks	Unified, interpretable, causal	High complexity	First holistic framework

discusses recent state-of-the-art methods using temporal GNN and graph evolution models applied to legal data and to other data fields.

Recent studies have increasingly explored multi-scale graph interaction [17], collaborative spatial-channel representations [18], and deep transfer learning [19] to improve the extraction and transfer of complex relational patterns. Such approaches demonstrate the value of combining multi-level structural representations with transferable knowledge, providing methodological motivation for models that must operate on evolving relational systems. Similarly, machine-vision-driven physics-informed neural networks [20] integrate visual observations with domain constraints, illustrating how data-driven representations can be strengthened through explicit structural or knowledge-based constraints. These developments are conceptually related to the present work because predictive jurisprudence also requires the integration of heterogeneous evidence rather than relying exclusively on isolated textual or graph representations.

Recent work has also investigated fault-tolerant and adaptive intelligent control under disturbances and system failures. For example, the study in [21] investigates fixed-time adaptive smooth control with relaxed performance constraints in the presence of aerodynamic interference and actuator faults. Such research highlights the importance of maintaining reliable decision performance under uncertainty, disturbances, and partial system degradation.

Across these works, strong individual capabilities (structure, semantics, or temporal modeling) emerge. But no unified, explainable, and causal framework exists. This gap justifies the necessity and novelty of the proposed predictive jurisprudence framework.

Further surveys [22], studies [23,24,25], theses [26,27], books [28], and chapters [29] investigate the integration of graph neural networks, causal inference, explainable AI, and advanced legal analytics for modeling complex legal reasoning and evolving jurisprudence. The following section further discusses this research gap.

2.2. Critical Discussion and Research Gap

Despite rapid progress in graph-based and AI-driven modeling, existing state-of-the-art approaches exhibit three fundamental limitations:

-

Fragmentation across modeling dimensions: Most works focus on one aspect only: TGNN-based models capturing structural evolution but ignoring semantics and causality, NLP models (such as BERT) capturing semantic understanding but ignoring graph structure, or causal models focusing on cause–effect relationships but lacking temporal graph context. Indeed, even advanced hybrid approaches (such as temporal heterogeneous GNNs) only partially integrate structure and semantics, leaving out explainability and causal reasoning.

-

The scarcity of predictive models adopted in law: Most current legal AI research, such as CaseGNN, focuses on retrieval, classification, and similarity but does not define or quantify precedent deviations, nor does it provide predictive tools for legal change.

-

Lack of intelligent platforms based on explainable causal law: Legal systems rely on principles of reasoning and traceable decisions. Current models (such as deep models and TGNNs) are black boxes that provide no explanation for the decisions made. Furthermore, these models do not rely on causal reasoning with graph learning. This gap is significant because legal AI needs to achieve the highest levels of transparency and accountability.

Furthermore, most modern studies rely on experiments without formal guarantees such as reliability and stability. In addition, recent surveys raise the issue of the theoretical foundations in graph learning.

The introduced system aims to find solutions to these gaps by combining structure, semantics, and causality into a single model, providing a formal definition of precedent deviation and enabling the prediction of legal development, not just analysis, with the possibility of interpretation.

As a conclusion, it is the conceptual incompleteness, not the performance, which represents the main issue of current legal systems. No existing approach simultaneously answers the following questions: How do legal systems evolve structurally? Why do they evolve (causality)? What concepts drive the evolution (semantics)? How does one explain and trust predictions?

This work is the first to address all four dimensions, establishing predictive jurisprudence as a new, unified research paradigm.

3. Methods and Material

This section highlights the mathematical modeling of the proposed framework, its algorithmic architecture, the proposed TGNN, DTM, the introduced theorems, and the explainable GNNExplainer model.

3.1. Problem Formulation and Preliminaries

We model an evolving legal system as a temporal attributed directed graph sequence:

G_1:T = {G_t}_t=1^T,G_t = (V_t,E_t,X_t,A_t)

(1)

where V_t is the set of cases decided up to time t, E_t ⊆ V_t × V_t are directed citations (edge j  →  i indicates case i cites j), A_t is the adjacency matrix, and X_t ∈ R^∣Vt∣×d are node features derived from texts (opinions, headnotes, statutes referenced). Each node i has timestamp t_i ≤ t, document D_i, and latent doctrinal state z_i^t.

We define a prediction target, Y_i^t, capturing doctrinal outcome (e.g., issue classification, overruling risk, or influence score). The task is to learn a mapping

$${\mathrm{f}}_{\mathsf{\Theta }}:\hspace{0.25em}({\mathrm{G}}_{1:\mathrm{t}},{\mathrm{X}}_{1:\mathrm{t}}) \mapsto {\widehat{\mathrm{Y}}}_{\mathrm{t}+1:\mathrm{t}+\mathsf{\tau }}, {\widehat{\Delta }}_{\mathrm{t}+1:\mathrm{t}+\mathsf{\tau }},$$

(2)

that forecasts both future outcomes and precedent drift over horizon τ.

The precedent drift is formulated as follows:

Let

Δ_t = α Δ_t^struct + β Δ_t^sem + γ Δ_t^causal,

(3)

where the structural, semantic, and causal components quantify changes in citation structure, topic distributions, and estimated causal influences, respectively. The parameters ($\alpha$, $\beta$, $\gamma$) are set to (0.4, 0.3, 0.3), consistent with the experimental configuration.

For the binary drift-prediction experiment, a future precedent is labeled drift-positive ($Y^{\mathrm{d}\mathrm{r}\mathrm{i}\mathrm{f}\mathrm{t}}$ = 1) when its future drift score exceeds the predefined drift threshold, and drift-negative ($Y^{\mathrm{d}\mathrm{r}\mathrm{i}\mathrm{f}\mathrm{t}}$ = 0) otherwise. Importantly, the threshold is determined using the training portion of each temporal split and is not estimated from future test data, thereby preventing temporal information leakage. With

(i)

Structural drift Δ_t^struct = ‖A_t − A_t−1‖_F or spectral distances;

(ii)

Semantic drift Δ_t^sem = KL(β_t‖β_t−1) over topic parameters; and

(iii)

Causal drift Δ_t^causal = ∑_(j,i)∣CE_j→i^t − CE_j→i^t−1∣.

The following assumptions are considered:

(A1)

Causal acyclicity in the contemporaneous graph.

(A2)

Temporal Markovity of order 1.

(A3)

Smooth semantic evolution: ‖β_t − β_t−1‖ ≤ L_β.

(A4)

Bounded degree/weights.

In the suggested formulation, the precedents are considered as edges, and legal language as features. Combined with time, these aspects determine the shifting of jurisprudence.

3.2. Algorithmic Pipeline

Figure 1 highlights the proposed jurisprudence framework and how to extract the predictions and explanations from the legal context.

Figure 2 shows the architecture of the jurisprudence framework. This architecture involves the judicial phases used to achieve actionable forecasts.

The complexity of the introduced framework is as follows:

-

Complexity of TGNN: O(∑_t∣E_t∣d).

-

Complexity of DTM: O(TKV).

-

Complexity of causal inference: O(∣E∣log n) optimized via sampling.

Hence, the model is scalable even when using large-size datasets (such as Free Law Project and LePaRD)

3.3. Temporal Graph Neural Network (TGNN)

We adopt a continuous-time TGNN with attention:

h_i^t = σ (W_xx_i^t + ∑_j∈Ni(t)  α_ij^t W_hh_j^t−), α_ij^t = exp(ϕ(h_i^t−, h_j^t−, Δt))/∑_kexp(ϕ(⋅))

(4)

We use temporal encoding in ϕ and residual connections for stability.

Loss is modeled as L_TGNN = L_link + L_node with negative sampling.

From a legal perspective, TGNN captures how influential cases propagate through citations over time, modeling the evolving backbone of precedent.

3.4. Dynamic Topic Modeling (DTM)

The DTM is modeled as follows:

For each document D_i at time t:

θ_i^t = LN(μ_t, Σ_t), w = Mult(θ_i^tβ_t)

(5)

with state evolution β_t = β_t−1 + ϵ_t, ϵ_t∼N(0,σ²I).

The objective is to model a variational ELBO with temporal smoothness regularization.

The semantic drift is:

Δ_t^sem = KL(β_t‖β_t−1)

(6)

DTM tracks the rise and decline of doctrines (e.g., standards, tests), linking language shifts to jurisprudential change.

3.5. Causal Inference on Precedent Graphs

To avoid ambiguity between classification, forecasting, and causal estimation, the proposed framework defines three distinct prediction targets. Let G_t = (V_t,E_t) denote the temporal citation graph at time t, X_t denote the textual/semantic representation of the legal cases, and D_t represent the proposed precedent-drift representation. The primary classification target is the binary variable

$${\mathrm{Y}}_{\mathrm{i}}^{ \mathrm{d}\mathrm{r}\mathrm{i}\mathrm{f}\mathrm{t}}(\mathrm{t}+\mathsf{\tau })\in \left\{\mathrm{0,1}\right\}$$

(7)

where ${\mathrm{Y}}_{\mathrm{i}}^{ \mathrm{d}\mathrm{r}\mathrm{i}\mathrm{f}\mathrm{t}}(t+\mathsf{\tau })=1$ indicates that precedent i exhibits a meaningful doctrinal drift during the future prediction window [t, t+$\mathsf{\tau }$], and ${\mathrm{Y}}_{\mathrm{i}}^{ \mathrm{d}\mathrm{r}\mathrm{i}\mathrm{f}\mathrm{t}}(t+\mathsf{\tau })=0$ otherwise.

The second target is future citation-link prediction,

$${\mathrm{Y}}_{\mathrm{i}\mathrm{j}}^{ \mathrm{c}\mathrm{i}\mathrm{t}\mathrm{e}}(\mathrm{t}+\mathsf{\tau })\in \left\{\mathrm{0,1}\right\}$$

(8)

where the value is one when a citation/influence edge between cases i and j is observed in the subsequent temporal window and zero otherwise. The third target is continuous causal-influence estimation,

$${CE }_{j\to i}^{\mathrm{t}+\mathsf{\tau }}\in \mathrm{R}$$

(9)

which represents the estimated causal influence of precedent j on precedent i. Unlike the first two targets, this is not treated as a binary classification problem and is therefore evaluated using causal-effect estimation error.

Accordingly, accuracy, precision, recall, and F1-score are reported only for the binary classification tasks, whereas ROC-AUC is reported for the corresponding binary prediction tasks and MSE is used for continuous drift and causal-effect estimation.

The causal effect is

CE_j→i^t = E[Y_i^t ∣ (A_ji = 1)] − E[Y_i^t ∣ (A_ji = 0)]

(10)

We identify effects via backdoor adjustment sets Z_ij^t (e.g., prior citations, court level, topics) and estimate using doubly robust learners. This separates genuine precedent influence (binding/persuasive authority) from coincidental co-citation or topical similarity.

3.6. Joint Learning Objective

Joint optimization ensures structural, semantic, and causal signals coherently reflect legal evolution rather than isolated proxies.

L = L_TGNN + λ₁L_DTM + λ₂L_causal + λ₃L_drift + λ₄L_align

(11)

where L_align = ∑_I ‖h_i^t − Pθ_i^t‖² aligns graph and topic spaces.

3.7. Theoretical Analysis

We propose two theorems to model the identifiability and stability of the precedent drift estimation.

Theorem 1. 

Identifiability of precedent drift. Under (A1)–(A4), if for every edge (j, i) there exists a valid backdoor set Z_ij^t blocking all backdoor paths from A_ji to Y_i^t, and the positivity condition holds, then {CE_j→i^t}_(j,i),t are identifiable. If, additionally, the TGNN encoder is injective over (A_t, X_t) up to graph isomorphism and DTM parameters β_t are uniquely determined (up to label permutation), then Δ_t is identifiable up to topic label permutation.

Proof of Theorem 1. 

• Causal part: By Pearl’s backdoor criterion, for each (j, i, t),

  E[Y_i^t ∣ do(A_ji = a)] = ∑_zE[Y_i^t ∣ A_ji = a, Z_ij^t = z] P(Z_ij^t = z),

  (12)
  • which is estimable from observational data given positivity and correct Z_ij^t.
  • Hence CE_j→i^t is identifiable.
• Structural part: Distances Δ_t^struct are functions of A_t and A_t−1. Since A_t is observed, identifiability is trivial. When using learned embeddings, the injectivity of the TGNN encoder ensures a one-to-one mapping (modulo isomorphism), so distances computed in the embedding space correspond to those in the graph space.
• Semantic part: In DTM, the parameters β_t are identifiable up to permutation under standard conditions (non-degeneracy and sufficient separation). The KL divergence is invariant to label permutation when consistently aligned; hence, Δ_t^sem is identifiable.
• Aggregation: Since each component is identifiable (up to permissible equivalences), their weighted sum Δ_t is identifiable up to topic label permutation. □

Theorem 2. 

Stability of drift estimation. Assume the TGNN update F is Lipschitz in its inputs with constant L_F < 1, i.e.,

‖F(H, A, X) − F(H′,A′,X′)‖ ≤ L_F(‖H − H′‖ + ‖A − A′‖ + ‖X − X′‖).

(13)

If ‖A_t − A_t−1‖ ≤ δ_A and ‖X_t − X_t−1‖ ≤ δ_X, then the embedding sequence satisfies

‖H_t − H_t−1‖ ≤ (L_F/(1 − L_F)) (δ_A + δ_X),

(14)

and the drift estimator $\widehat{\Delta }$_t is uniformly stable, i.e., perturbations in inputs induce bounded changes in $\widehat{\Delta }$_t.

Proof of Theorem 2. 

Let

H_t = F(H_t−1, A_t, X_t) and H_t−1 = F(H_t−2, A_t−1, X_t−1).

(15)

Then,

‖H_t − H_t−1‖ ≤ L_F (‖H_t−1 − H_t−2‖ + ‖A_t − A_t−1‖ + ‖X_t − X^t−1‖)

(16)

Unrolling the recursion yields a geometric series:

$$\Vert {\mathrm{H}}_{\mathrm{t}} - {\mathrm{H}}_{\mathrm{t}-1}\Vert \le {\mathrm{L}}_{\mathrm{F}}{\sum }_{k=0}^{\mathrm{\infty }}{L}_{\mathrm{F}}^{\mathrm{k}}({\mathsf{\delta }}_{\mathrm{A}} + {\mathsf{\delta }}_{\mathrm{X}}) = \left({\displaystyle \frac{\mathrm{L}\mathrm{F}}{1-\mathrm{L}\mathrm{F}}}\right) ({\mathsf{\delta }}_{\mathrm{A}} + {\mathsf{\delta }}_{\mathrm{X}}).$$

(17)

Since $\widehat{\Delta }$_t is Lipschitz in (H_t, H_t−1, β_t, β_t−1, CE^t, CE^t−1) and each term has bounded variation (DTM smoothness and bounded CE changes), the overall estimator is uniformly stable. □

Stability ensures that small updates in case law or language do not cause erratic predictions, crucial for reliable legal decision support.

3.8. Explainability via GNNExplainer

Given a trained TGNN and a target node i, GNNExplainer learns a mask over edges and features:

$${max}_{ME,MX} \mathrm{I}({\mathrm{Y}}_{\mathrm{i}}; \mathrm{G}\odot {\mathrm{M}}_{\mathrm{E}}, \mathrm{X}\odot {\mathrm{M}}_{\mathrm{X}}) - {\lambda }_{\mathrm{E}}{\Vert \mathrm{ME}\Vert }_1 - {\lambda }_{\mathrm{X}}{\Vert {\mathrm{M}}_{\mathrm{X}}\Vert }_1$$

(18)

To optimize the explanation of deep decisions, continuous relaxations with sigmoid masks and gradient descent were employed. The outputs are represented by influential textual features and the salient precedent subgraph G_S. The predictions are directed by the explanations via the selection of prior cases and doctrines and via enabling the auditability of judicial reasoning.

4. Results

The experiments rely on two datasets: Free Law Project [30] and LePaRD [31]. These datasets are large and involve real data. A time-division approach is applied to data from the two datasets to conduct experiments on TGNN, DTM, and causal module models. The following metrics are considered in conducting experiments: ROC curve analysis, attentional perception, robustness under noise conditions, long-term predictive power, generalizability across jurisdictions, time skew evolution, computational efficiency, interpretability, and statistical significance.

4.1. Baseline Selection and Fair Experimental Protocol

To ensure a fair and reproducible comparison, the baseline models were selected to represent complementary families of approaches relevant to predictive jurisprudence. GCN and GAT represent conventional static graph-learning approaches, whereas TGN represents temporal graph learning. BERT and LDA + SVM provide text-based semantic baselines, while DynamicTriad represents dynamic graph embedding. CausalForest provides a causal-inference baseline without joint temporal graph and semantic modeling. In addition, CaseGNN is considered a legal graph-text baseline because it combines textual information with citation structure. CaseGNN is static and was designed for retrieval rather than temporal legal prediction. In addition, CaseGNN was developed for legal case retrieval using text-attributed case graphs. Instead of the standard CaseGNN, we propose CaseGNN-adapted as a legal-domain baseline because the original CaseGNN task is legal case retrieval rather than future doctrinal-drift classification. In CaseGNN-adapted, the CaseGNN architecture is adapted to the prediction task by replacing its retrieval objective with the same binary drift-prediction output layer used by the other baselines.

All models use exactly the same FLP and LePaRD temporal partitions. The temporal protocol preserves the original principle of training on historical observations and evaluating on subsequent observations, avoiding random mixing of future cases into the training set. All models receive the same normalized citation graph, temporal snapshots, case identifiers, and standardized textual representations whenever these inputs are required by the corresponding model. The preprocessing pipeline already includes temporal slicing, tokenization, transformer-based embeddings, citation normalization, and removal of nodes with fewer than three degrees. The same predefined hyperparameter-search budget and validation folds are used for all trainable baselines. The selected parameters are reported together with the proposed model’s settings. All iterative learning models use the same maximum training budget and an identical validation-based early-stopping criterion, with the best validation checkpoint retained. All methods are evaluated using the same folds and metrics and, where applicable, drift-prediction error and causal-effect estimation error.

Transformer-based textual embeddings were precomputed once and stored. When precomputed, the embeddings were generated once for the retained legal documents and subsequently loaded as node features during TGNN training. Consequently, the reported GPU training time does not include repeated transformer encoding of the complete raw corpus during every training epoch.

To avoid circular evaluation, the target variables used for predictive assessment are constructed independently of the proposed model. Model-generated quantities, including DTM topic trajectories, TGNN attention weights, GNNExplainer outputs, and estimated causal effects, are not used to construct the evaluation labels. For future doctrinal-drift prediction, the reference outcome is derived from independently observable legal developments occurring after the prediction time. The reference event is defined according to the legal-event annotation available in the evaluation corpus and includes documented changes in the treatment or status of a precedent. The model is trained exclusively on information available before the prediction cutoff, while the reference event is determined from subsequent legal records. For causal-effect analysis, the causal module produces an estimated treatment effect under the stated causal assumptions. Its validity is evaluated using independent legal outcomes and sensitivity/robustness analyses rather than by comparing the estimate against a model-generated causal target.

4.2. Experimental Setup

To test the proposed model against other paradigms, we conduct experiments on two datasets: the first one is a large-scale legal corpus with citation graphs named Free Law Project (FLP) [30], and the second dataset is a benchmark for semantic legal retrieval, named LePaRD [31].

A phase of preprocessing is achieved on these data to guarantee an accurate representation of precedent relationships and evolving legal semantics: First, we establish a citation graph using directed and temporal edges. Then, temporal slicing (yearly snapshots), tokenization and transformer embeddings are established. Afterward, normalization of legal citations is accomplished. Finally, nodes with less than three degrees are filtered.

The validation protocol relies on 5-fold cross-validation. This temporal decomposition allows training on past data and testing on future data. Moreover, average standard deviations are used on all metrics.

The study uses the U.S. case-law corpus provided through Free Law Project/CourtListener [30]. The raw corpus is subsequently filtered and transformed into a temporal citation graph. Accordingly, the statistics reported for the experiments refer to this fixed snapshot and derived subset rather than to the continuously changing online database.

We use LePaRD [31], a large-scale dataset of U.S. federal judicial citations to precedent in context. The original resource contains approximately 1.8 million unique target passages associated with approximately 4.3 million citation contexts. It is designed primarily for legal passage retrieval rather than as a conventional case-level citation network. In the present study, LePaRD is transformed into a temporal legal-representation graph by linking citation contexts to their corresponding precedent passages and associating each observation with its publication date and court information. Therefore, the number of graph nodes and edges used in our experiments should not be interpreted as the original dataset statistics. The experiments use the 10K most-cited-passage version of LePaRD, rather than the complete all-data release.

The datasets used contain different temporal citation structures and legal semantics, which allows precedent drift modeling and assessment of predictive jurisprudence under realistic judicial conditions.

For reproducibility, all experiments were conducted using an NVIDIA GeForce RTX 4090 GPU with 24 GB of VRAM, an AMD Ryzen 9 7950X 16-core CPU, 64 GB of system RAM, and a 2 TB NVMe SSD. The software environment consisted of Python 3.10.14, PyTorch 2.2.2, PyTorch Geometric 2.5.2, and CUDA 12.1. The training was conducted using a batch size of 64 and 100 training epochs, with the same hardware and software configuration used for the proposed framework and all baseline models.

Table 2. Parameters of the suggested system.

Parameter	Value
TGNN layers	3
Hidden dimension	128
Attention heads	4
Learning rate	0.001
Batch size	64
Epochs	100
Topic number (DTM)	50
Drift weights (α, β, γ)	(0.4, 0.3, 0.3)
Dropout	0.3

highlights the parameters of the suggested TGNN-DTM-XAI causal-based system, while

Table 3. Hyperparameters of the baseline models used.

Model	Parameter	Value
GCN	Layers	2
	Hidden dim	128
	LR	0.01
	Dropout	0.5
GAT	Layers	2
	Heads	8
	Hidden dim	128
	LR	0.005
TGN	Memory dim	128
	Time encoding	Sinusoidal
	LR	0.001
DynamicTriad	Embedding dim	128
	Window size	5
LDA + SVM	Topics	50
	SVM kernel	RBF
	C	1.0
BERT	BERT Model	base (768 dim)
	LR	2 × 10−5
	Epochs	5
CausalForest	Trees	500
	Max depth	10

illustrates the parameters of the standard systems. All indicated values in both tables are set after extensive empirical tests.

4.3. Quantitative Performance

Table 4. Binary precedent-drift classification performance on FLP and LePaRD.

	Accuracy (%)		F1-Score		Precision		Recall
Model	FLP	LePaRD	FLP	LePaRD	FLP	LePaRD	FLP	LePaRD
GCN	78.2 ± 1.2	76.5 ± 1.4	0.76 ± 0.01	0.74 ± 0.02	0.75 ± 0.01	0.73 ± 0.02	0.76 ± 0.01	0.74 ± 0.02
GAT	80.5 ± 1.1	78.9 ± 1.3	0.79 ± 0.01	0.77 ± 0.01	0.78 ± 0.01	0.76 ± 0.02	0.79 ± 0.01	0.78 ± 0.01
TGN	84.7 ± 0.9	82.6 ± 1.0	0.83 ± 0.01	0.81 ± 0.01	0.82 ± 0.01	0.80 ± 0.01	0.83 ± 0.01	0.82 ± 0.01
BERT	82.3 ± 1.0	83.5 ± 0.8	0.81 ± 0.01	0.82 ± 0.01	0.80 ± 0.01	0.82 ± 0.01	0.81 ± 0.01	0.83 ± 0.01
LDA + SVM	75.1 ± 1.5	74.3 ± 1.6	0.73 ± 0.02	0.72 ± 0.02	0.72 ± 0.02	0.71 ± 0.02	0.73 ± 0.02	0.72 ± 0.02
CaseGNN-adapted	83.5 ± 0.9	81.3 ± 1.2	0.76 ± 0.03	0.78 ± 0.01	0.81 ± 0.01	0.77 ± 0.02	0.82 ± 0.01	0.80 ± 0.02
Proposed	89.6 ± 0.7	87.9 ± 0.8	0.88 ± 0.01	0.86 ± 0.01	0.88 ± 0.01	0.85 ± 0.01	0.88 ± 0.01	0.87 ± 0.01

compares the predictive performance of the proposed framework against state-of-the-art baselines using multiple evaluation metrics.

Table 4 demonstrates the performance of the proposed framework on the binary future precedent-drift classification task. Table 4 highlights the superiority in performance of the suggested model across all metrics, which indicates the value of combining causal inference and semantic evolution with temporal graph learning.

To further assess the proposed model on legal prediction tasks, we test it on the Receiver Operating Characteristic (ROC) metric.

Table 5. ROC-AUC for citation-link prediction, precedent-drift classification, and causal-influence prediction.

Model	Citation Prediction AUC	Drift Prediction AUC	Causal Influence AUC	Mean ROC-AUC
GCN	0.81	0.77	0.73	0.77
GAT	0.84	0.80	0.76	0.80
TGN	0.88	0.85	0.81	0.85
BERT	0.86	0.79	0.75	0.80
CaseGNN-adapted	0.89	0.89	0.83	0.87
Proposed	0.93	0.91	0.89	0.91

compares the ROC-AUC values obtained by our model with other standard approaches such as GAT, TGN and BERT.

Table 5 confirms that the suggested model has the best (highest) ROC-AUC values across all prediction tasks. Moreover, once again, the high values of causal influence AUC indicate the effectiveness of combining causal inference with temporal graph learning, on detecting the difference between stable precedential structures and evolving legal patterns.

4.4. Drift Prediction Analysis

To assess the performance of our jurisprudence model in predicting precedent drift on dynamic evolving legal systems, we compute prediction errors and causal-influence estimation. The findings, shown in

Table 6. External-evidence-based evaluation of future doctrinal drift.

Model	FLP MSE	LePaRD MSE
TGN	0.084 ± 0.004	0.089 ± 0.005
DynamicTriad	0.091 ± 0.006	0.095 ± 0.006
LDA-based	0.110 ± 0.007	0.115 ± 0.008
Proposed	0.052 ± 0.003	0.058 ± 0.004

, show how our model identifies causal, semantic, and structural evolutions that shape the jurisprudential development over time.

Because the composite drift score in Equation (3) is generated by the proposed framework, it is not used as an independent ground truth for model comparison. Instead, future doctrinal change is evaluated against independently observable legal outcomes available after the prediction time. These outcomes are constructed without using the proposed model’s DTM representation, attention weights, or causal estimates. A prediction is considered correct when the model identifies a precedent associated with an independently documented doctrinal-change event within the predefined future window. This protocol prevents the proposed model from being evaluated against targets generated by its own components.

Table 6 indicates that the proposed model significantly reduces drift prediction error, confirming the effectiveness of combining structural, semantic, and causal signals. Unlike Table 4, which evaluates binary drift classification, Table 6 evaluates the continuous future drift score. Mean squared error (MSE) is therefore used to quantify the difference between predicted and observed drift scores.

The explicit causal integration shown in

Table 7. Observational causal-effect estimation and external outcome validation.

Model	FLP Error	LePaRD Error
CausalForest	0.072 ± 0.005	0.075 ± 0.006
TGN (no causal)	0.095 ± 0.006	0.099 ± 0.007
Proposed	0.041 ± 0.003	0.045 ± 0.004

leads to substantially more accurate estimation of precedent influence. Table 7 evaluates the continuous estimation of causal precedent influence. The reported error therefore measures the deviation between the estimated and reference causal-effect values and should not be interpreted as classification accuracy.

Table 8. Relative contribution of structural, semantic, and causal components to the continuous precedent-drift score.

Drift Component	FLP Contribution (%)	LePaRD Contribution (%)
Structural Drift	41.5	39.8
Semantic Drift	31.2	34.5
Causal Drift	27.3	25.7

analyzes the relative contribution of structural, semantic, and causal components in overall precedent drift estimation.

The component contribution analysis is interpreted as an internal decomposition of the proposed drift representation, rather than as evidence that structural, semantic, or causal factors objectively account for a corresponding percentage of real-world doctrinal change. Because the components are generated within the proposed framework, these values are useful for understanding the model’s representation but do not constitute independent causal attribution.

Structural drift contributes most significantly to legal evolution, indicating that citation dynamics remain the dominant signal in jurisprudential change. However, semantic and causal components substantially improve prediction accuracy and interpretability.

4.5. Temporal Drift and Temporal Topic Evolution Analysis

To study the manner in which jurisprudence evolves over time, our model is used to quantify precedent drift on numerous temporal epochs as shown in

Table 9. Temporal drift evolution analysis.

Period	Structural Drift	Semantic Drift	Causal Drift	Overall Drift Score
2000–2005	0.21	0.18	0.14	0.18
2005–2010	0.29	0.24	0.19	0.23
2010–2015	0.37	0.31	0.25	0.30
2015–2020	0.45	0.39	0.34	0.41
2020–2025	0.52	0.47	0.41	0.46

.

Table 9 demonstrates a progressive increase in precedent drift, which indicates that current legal systems evolve quickly because of recent fast changes in society, in technological advances, and in legal complexity. Structural drift remains the leading element, replicating the changing aspect of citation networks, whereas the growth of semantic and causal drift reflects emergent doctrinal transformation and variations in legal influence patterns.

The evolution of major legal doctrines over time allows further weighing of the DTM module.

Table 10. Temporal topic evolution analysis.

Legal Topic	2000–2005	2005–2010	2010–2015	2015–2020	2020–2025
Digital Privacy	0.05	0.09	0.18	0.31	0.44
Cybercrime	0.04	0.08	0.16	0.28	0.39
Constitutional Rights	0.26	0.28	0.31	0.34	0.36
Corporate Liability	0.18	0.21	0.25	0.27	0.29
Environmental Law	0.09	0.11	0.17	0.24	0.33

shows the evolution of different relevant legal topics from 2000 to 2025 using the data from the Free Law Project corpus [30].

The outcomes in Table 10 reveal that evolving legal fields such as digital privacy and cybercrime experience rapid semantic evolution over time. These results then confirm the ability of DTM to identify doctrinal transitions.

4.6. Backtesting Historical Legal Shifts

In this set of experiments, we study how historical legal data affect the performance of our prediction model by comparing the resulting predictions with the evolved legal patterns in the data. These comparisons show the level of participation of our framework in the continuous jurisprudential shifts.

Table 11. Backtesting accuracy.

Event Type	FLP Accuracy (%) (Mean ± SD)	FLP 95% CI	LePaRD Accuracy (%) (Mean ± SD)	LePaRD 95% CI
Major shifts	91.2 ± 0.8	[90.5, 91.9]	89.5 ± 1.0	[88.6, 90.4]
Moderate shifts	87.5 ± 1.0	[86.6, 88.4]	85.9 ± 1.2	[84.8, 87.0]
Minor changes	82.8 ± 1.3	[81.6, 84.0]	81.4 ± 1.4	[80.1, 82.7]

reveals the efficiency of the framework in predicting the main doctrinal changes, which demonstrates a better structural and semantic predictive ability.

For further temporal experiments,

Table 12. Long-term forecasting performance.

Forecast Horizon	FLP Accuracy (%) (Mean ± SD)	LePaRD Accuracy (%) (Mean ± SD)
1 Year	89.6	87.9
3 Years	86.4	84.8
5 Years	82.7	80.9
10 Years	75.8	73.2

presents the results of long-term forecasting performance on extended prediction horizons.

Prediction performance gradually decreases as the forecasting horizon increases, which is expected due to growing uncertainty in long-term legal evolution. However, the results of the model indicate good performance, even for extended periods.

Historical events were selected from the Free Law Project corpus based on three criteria: (i) the event represented a documented judicial decision associated with a recognizable doctrinal development; (ii) sufficient pre-event and post-event citation/textual information was available to construct temporal observations; and (iii) the event generated observable changes in the semantic, structural, and/or causal characteristics of the relevant precedent network.

The event-level results are summarized in

Table 13. Representative historical jurisprudential events used for backtesting.

No.	Historical Event	Year	Legal Topic	Doctrinal Development	Shift Category	Predicted Drift	Event Detected
1	United States v. Jones	2012	Digital privacy	GPS tracking recognized as a Fourth Amendment search	Major	0.79	Yes
2	Riley v. California	2014	Digital privacy	Warrant generally required for search of digital contents of a seized mobile phone	Major	0.86	Yes
3	Obergefell v. Hodges	2015	Constitutional rights	Recognition of constitutional right to same-sex marriage	Major	0.91	Yes
4	Carpenter v. United States	2018	Digital privacy	Fourth Amendment protection extended to historical cell-site location information	Major	0.94	Yes
5	Bostock v. Clayton County	2020	Constitutional rights	Title VII interpreted to prohibit discrimination based on sexual orientation or gender identity	Major	0.88	Yes
6	Digital-search warrant cases following Riley	2015–2017	Digital privacy	Incremental clarification of digital-search doctrine	Moderate	0.71	Yes
7	Post-Carpenter lower-court applications	2019–2021	Digital privacy	Extension/refinement of privacy principles to emerging digital records	Moderate	0.69	Yes
8	Incremental environmental-law precedent developments	2018–2020	Environmental law	Narrow doctrinal refinements	Minor	0.48	Yes

. The backtesting procedure used only information available up to each historical prediction point. The model generated a predicted drift score and forecast direction for the subsequent temporal window. The prediction was then compared with the documented post-event jurisprudential development.

Each event was categorized according to the magnitude of the resulting doctrinal change. Major shifts were defined as decisions establishing a substantially new doctrinal rule, materially limiting an established rule, or producing a substantial subsequent change in precedent structure and legal semantics. Moderate shifts were defined as substantial extensions, refinements, or reinterpretations of an existing doctrine without fundamentally replacing the governing rule. Minor changes were defined as narrower applications, incremental modifications, or fact-specific developments producing comparatively limited downstream changes.

Qualitative Case Study: Evolution of Digital-Privacy Jurisprudence

To provide an interpretable example of the backtesting procedure, we examined the evolution of U.S. digital-privacy jurisprudence. This topic was selected because the DTM analysis in Table 10 shows a pronounced increase in its semantic prevalence, from 0.18 during 2000–2005 to 0.47 during 2020–2025, representing the strongest temporal growth among the legal topics considered.

The evolution provides a particularly suitable test of the proposed precedent-drift framework because the relevant jurisprudence developed through a sequence of temporally connected decisions addressing increasingly pervasive digital technologies. Three landmark decisions were therefore selected as representative events: United States v. Jones (2012), Riley v. California (2014), and Carpenter v. United States (2018).

In Jones, the Supreme Court considered the government’s use of a GPS tracking device attached to a vehicle. The decision represented an important development in Fourth Amendment jurisprudence concerning technologically enabled location surveillance. The model’s simulated pre-event drift score of 0.79 indicates a substantial increase in the combined structural, semantic, and causal drift components before the documented doctrinal transition.

The subsequent Riley decision provided a second major transition. The Supreme Court held that police generally require a warrant before searching the digital contents of a mobile phone seized from an arrested individual. The Court emphasized the qualitatively different quantity and nature of information contained in modern mobile devices. The model correspondingly produced a higher simulated drift score of 0.86, indicating that the digital-privacy trajectory had become increasingly pronounced.

The strongest transition occurred around Carpenter v. United States (2018). The Supreme Court held that an individual maintains a legitimate expectation of privacy in historical cell-site location information and rejected a simple application of the traditional third-party doctrine to this category of digital location data. The simulated drift score increased to 0.94, representing the largest predicted transition in this sequence.

The temporal ordering of these observations is important. The model does not use the later judicial outcome to infer the earlier prediction. Instead, the model is trained on information available before the prediction point and generates the anticipated drift for the subsequent period. The subsequent judicial decision is then used only as an external historical reference for evaluating whether the predicted transition corresponds to an observed doctrinal development.

Overall, this case study illustrates how semantic evolution, citation-network restructuring, and estimated causal influence can jointly identify an emerging doctrinal transition.

4.7. Attention Heatmap Analysis

To better understand the internal reasoning of the TGNN algorithm, the average attention distribution across different categories of legal data is computed.

Table 14. Analysis of attention weights of legal features.

Legal Feature Category	Average Attention Weight
Supreme Court precedents	0.34
Constitutional references	0.27
Statutory interpretation	0.18
Procedural citations	0.11
Secondary legal doctrines	0.10

presents the temporal attention mechanism and the associated weights.

The attention distribution shown in Table 14 indicates that the model prioritizes highly authoritative legal precedents and constitutional reasoning during prediction, which aligns with real judicial reasoning processes. This indicates that our model is able to identify meaningful precedent influence patterns and structures.

4.8. Evaluation of Explainability of Citation Subgraph

In this section, a citation subgraph analysis is employed to assess the reliability of explanations generated by GNNExplainer. As shown in

Table 15. Citation subgraph explainability evaluation.

Metric	FLP ± SD	FLP 95% CI	LePaRD ± SD	LePaRD 95% CI
Subgraph Fidelity	0.89 ± 0.02	[0.87, 0.91]	0.87 ± 0.01	[0.85, 0.89]
Explanation Compactness	0.81 ± 0.03	[0.78, 0.84]	0.79 ± 0.03	[0.76, 0.82]
Citation Relevance	0.92 ± 0.02	[0.90, 0.94]	0.90 ± 0.01	[0.88, 0.92]
Legal Consistency Score	0.88 ± 0.01	[0.86, 0.90]	0.86 ± 0.02	[0.83, 0.89]
Human Expert Agreement	0.85 ± 0.03	[0.82, 0.88]	0.83 ± 0.02	[0.80, 0.86]

, different metrics are involved, such as fidelity, consistency, compactness, and agreement with human experts. The resulting ICC is 0.85 for the evaluation associated with Table 15.

The citation subgraphs generated by the proposed framework achieve high fidelity and strong agreement with legal experts. The findings indicate that influential legal relationships and judicial reasoning structures are both well captured and represented by precedent chains.

4.9. Evaluation of Statistical Significance

To validate the reliability of the obtained performance improvements, statistical significance testing was performed using paired t-tests across all cross-validation folds.

Table 16. Statistical significance analysis.

Comparison	t-Statistic	p-Value	Significant	Mean Difference	95% CI	Effect Size
Proposed vs. GCN	7.42	<0.001	Yes	8.4%	[6.0%, 10.8%]	1.45
Proposed vs. GAT	6.85	<0.001	Yes	7.1%	[4.9%, 9.3%]	1.32
Proposed vs. TGN	5.91	<0.001	Yes	5.2%	[3.1%, 7.3%]	1.18
Proposed vs. BERT	6.27	<0.001	Yes	6.4%	[4.2%, 8.6%]	1.27

summarizes the obtained results.

The low p-values confirm that the performance improvements achieved by the proposed framework are statistically significant and not caused by random fluctuations.

4.10. Evaluation of Robustness

Robustness experiments were conducted by introducing controlled perturbations into citation structures and textual representations. The robustness of the system on noisy citation structures is tested in

Table 17. Robustness under noisy legal data.

Noise Level	FLP Accuracy (%)	95% CI	LePaRD Accuracy (%)	95% CI
0%	89.6	[89.0, 90.2]	87.9	[87.2, 88.6]
5%	88.2	[87.5, 88.9]	86.7	[86.0, 87.4]
10%	86.9	[86.2, 87.6]	85.1	[84.3, 85.9]
20%	83.8	[83.0, 84.6]	81.9	[81.0, 82.8]
30%	79.6	[79.3, 81.2]	77.5	[75.4, 78.3]

.

The proposed framework maintains relatively stable performance even under substantial noise levels, demonstrating strong robustness against imperfect legal data and citation inconsistencies. This robustness is due to the combination of structural, semantic, and causal components into one framework.

To further assess the practical robustness of the proposed framework, we set in

Table 18. Robustness of the proposed framework under data and distribution perturbations.

Robustness Condition	Level	Accuracy (%)	F1-Score (%)	ROC-AUC	Drift MSE	Causal Error
Clean data	0%	89.6 ± 0.7	89.5 ± 0.7	0.93	0.052 ± 0.003	0.041 ± 0.003
Input noise	5%	88.2 ± 0.8	88.1 ± 0.8	0.92	0.056 ± 0.003	0.043 ± 0.003
	10%	86.9 ± 0.9	86.7 ± 0.9	0.90	0.061 ± 0.004	0.046 ± 0.004
	20%	83.8 ± 1.1	83.5 ± 1.1	0.87	0.071 ± 0.005	0.052 ± 0.005
Missing citations	10%	88.1 ± 0.8	87.9 ± 0.8	0.91	0.059 ± 0.004	0.045 ± 0.004
	20%	85.9 ± 1.0	85.6 ± 1.0	0.89	0.068 ± 0.005	0.049 ± 0.004
	30%	82.7 ± 1.2	82.3 ± 1.2	0.85	0.079 ± 0.006	0.056 ± 0.005
Sparse history	75%	88.0 ± 0.8	87.8 ± 0.8	0.91	0.058 ± 0.004	0.044 ± 0.003
	50%	85.4 ± 1.0	85.1 ± 1.0	0.88	0.067 ± 0.005	0.049 ± 0.004
	25%	81.6 ± 1.3	81.2 ± 1.3	0.84	0.081 ± 0.006	0.057 ± 0.005
Temporal shift	+1 year	88.7 ± 0.8	88.5 ± 0.8	0.92	0.057 ± 0.004	0.043 ± 0.003
	+3 years	86.4 ± 1.0	86.1 ± 1.0	0.89	0.065 ± 0.005	0.048 ± 0.004
	+5 years	83.9 ± 1.2	83.6 ± 1.2	0.86	0.074 ± 0.006	0.054 ± 0.005
Jurisdiction shift	Cross-jurisdiction	84.8 ± 1.1	84.5 ± 1.1	0.87	0.072 ± 0.005	0.055 ± 0.005
Semantic shift	Emerging topics	82.9 ± 1.2	82.5 ± 1.2	0.85	0.078 ± 0.006	0.058 ± 0.005
Citation-structure shift	−20% edges	84.6 ± 1.1	84.2 ± 1.1	0.87	0.073 ± 0.005	0.054 ± 0.005

progressively more challenging conditions involving input noise, missing citation relationships, sparse historical evidence, temporal shift, jurisdictional shift, semantic shift, and changes in citation-network structure, while keeping the remaining experimental settings unchanged.

Overall, the results in Table 18 indicate a gradual rather than abrupt degradation as perturbation severity increases, with the largest reductions occurring under sparse historical evidence, substantial citation loss, long-term temporal shift, and emerging semantic topics.

4.11. Evaluation of Computational Efficiency

A GPU-based training environment was used to assess the computational efficiency of the different models against our framework. The findings are shown in

Table 19. Computational efficiency comparison.

Model	Training Time (h)	GPU Memory (GB)	Inference Time (ms/Sample)
GCN	4.1	6.2	11
GAT	6.5	8.1	17
TGN	11.3	12.4	25
Proposed Framework	15.8	14.9	31

. The reported 15.8 h training time and 14.9 GB GPU requirement are measurements for the processed experimental representation, not for simultaneously loading and training over the entire raw corpus.

Despite the addition of TGNN, DTM, causal inference, and explainability, which makes the introduced model computationally higher, the model outperforms other predictive models in terms of causal reasoning, interpretability, and accuracy.

The raw legal resources were not processed in their entirety as a single GPU-resident corpus. Instead, a reproducible preprocessing pipeline was applied. First, a fixed dataset snapshot was obtained. Second, documents outside the predefined temporal interval were removed. Third, legal citations were normalized and mapped to case identifiers or precedent-passage identifiers. Fourth, duplicate and unresolved citation records were removed. Fifth, temporal snapshots were constructed. Finally, nodes with insufficient graph connectivity were filtered before graph learning.

For each temporal snapshot, the resulting graph contains only the cases and citation relationships available up to that point. This representation is subsequently used to train the temporal graph models. Textual representations are generated separately and associated with the graph nodes rather than repeatedly encoding the complete raw corpus during every training epoch.

The reported GPU memory does not correspond to loading the complete raw legal corpus into GPU memory. Raw documents and citation records are stored on disk, while only the processed graph mini-batches and corresponding node representations required for the current training step are transferred to the GPU. Therefore, GPU memory consumption depends primarily on the sampled subgraph size, feature dimensionality, batch size, and model architecture rather than directly on the total disk size of the source corpus.

4.12. Evaluation of Cross-Jurisdiction Generalization

Table 20. Cross-jurisdiction generalization analysis.

Jurisdiction	Accuracy (%) ± SD	95% CI	Drift Prediction Error ± SD	95% CI
Federal Courts	89.6 ± 0.7	[88.9, 90.3]	0.052 ± 0.003	[0.049, 0.055]
State Courts	86.9 ± 0.9	[86.0, 87.8]	0.061 ± 0.004	[0.057, 0.065]
International Legal Cases	83.7 ± 1.1	[82.6, 84.8]	0.069 ± 0.005	[0.064, 0.074]
Administrative Courts	84.8 ± 1.0	[83.8, 85.8]	0.064 ± 0.004	[0.060, 0.068]

proposes a set of cross-jurisdiction tests to assess the accuracy and drift error of our model across different legal environments.

The proposed framework generalizes effectively across different judicial systems, although performance slightly decreases in highly heterogeneous legal environments due to variations in citation structures, procedural traditions, and doctrinal organization.

The main causes of prediction errors detected during experimentation are enumerated in

Table 21. Error categorization analysis.

Error Type	Percentage (%)
Ambiguous precedents	31.4
Sparse citation history	24.6
Semantic ambiguity	21.8
Temporal inconsistency	14.2
Annotation noise	8.0

.

Most of these errors stem either from fragmented historical citation structures or from obscure prior relationships. This demonstrates the considerable complexity of modeling legal reasoning patterns.

4.13. Evaluation of Human Experts

Table 22. Human expert scores for explainability.

Criterion	Judges Score ± SD	Legal Experts ± SD
Explanation clarity	4.5 ± 0.5	4.7 ± 0.4
Legal consistency	4.4 ± 0.6	4.8 ± 0.3
Trustworthiness	4.3 ± 0.6	4.6 ± 0.5
Practical usefulness	4.6 ± 0.5	4.7 ± 0.4

shows human expert evaluation of the generated explanations. Legal experts rated the explanations highly in terms of legal consistency and practical usefulness, indicating that the proposed framework produces explanations aligned with judicial reasoning processes. The number of participating experts was N = 8. The experts were asked to independently assess the relevance and practical usefulness of the explanations generated by the proposed framework. A five-point Likert scale was used, ranging from 1 (very poor/strongly disagree) to 5 (excellent/strongly agree). The experts evaluated the model explanations independently and were not provided with the ratings of other participants. The resulting ICC was 0.83 for the evaluation associated with

Table 23. Component analysis of the proposed framework.

Model Configuration	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	ROC-AUC	Drift MSE	Causal Error
TGNN only	84.9 ± 1.1	83.8 ± 1.2	85.7 ±1.1	84.7 ± 1.1	0.87	0.084 ± 0.006	0.061 ± 0.005
TGNN + DTM	86.9 ± 0.9	86.1 ± 1.0	87.6 ± 0.9	86.8 ± 0.9	0.89	0.069 ± 0.005	0.088 ± 0.006
TGNN + Causal Inference	87.8 ± 0.8	87.0 ± 0.9	88.5 ± 0.8	87.7 ± 0.8	0.90	0.063 ± 0.004	0.043 ± 0.003
TGNN + DTM + Causal inference	89.1 ± 0.7	88.5 ± 0.8	89.7 ± 0.7	89.0 ± 0.7	0.92	0.055 ± 0.003	0.042 ± 0.003
Complete framework (TGNN + DTM + Causal inference + GNNExplainer)	89.6 ± 0.7	88.9 ± 0.8	90.2 ± 0.7	89.5 ± 0.7	0.93	0.052 ± 0.003	0.041 ± 0.003

, which indicates a strong inter-rater consistency.

The expert evaluation is intended to provide external human validation of the interpretability and practical relevance of the generated explanations, rather than to establish definitive legal validity or causal correctness.

4.14. Component Contribution Analysis

The results in Table 23 indicate a progressive improvement as complementary information sources are incorporated. TGNN alone provides the structural and temporal baseline, while adding DTM improves the representation of semantic evolution. Adding causal inference further improves predictive performance and substantially reduces causal estimation error. The combination of TGNN, DTM, and causal inference produces further improvement, while the complete framework provides the best overall performance after incorporating alignment, forecasting, and explainability. This pattern is consistent with the framework’s design, where TGNN captures evolving citation structure, DTM models semantic drift, and causal inference distinguishes genuine precedent influence from correlation. Notably, GNNExplainer is not expected to produce a large change in predictive accuracy, because its primary role is interpretability rather than prediction.

The failure-case analysis connects the observed failure cases to the limitations of the proposed components. The achieved analysis identifies four major error sources: ambiguous precedents (31.4%), sparse citation history (24.6%), semantic ambiguity (21.8%), and temporal inconsistency (14.2%). This analysis demonstrates that the proposed framework should not be interpreted as uniformly effective for every legal prediction scenario. Sparse citation networks can limit TGNN-based structural reasoning, ambiguous language can reduce the reliability of DTM-derived semantic signals, and abrupt or inconsistent historical changes can challenge temporal extrapolation.

5. Discussion

The series of cross-validated experiments confirms the performance of the proposed predictive jurisprudence system. In general, this system was able to consider the structural and causal dimensions better than standard models due to the following characteristics:

-

Time-based modeling tools are in general effective in tracking legal evolution. Indeed, due to their incapacity in managing the dynamic behavior of case law, static paradigms, such as latent discrimination analysis and generative neural networks, fail in obtaining an accurate variance ratio. Nonetheless, Temporal GNN successfully identifies evolving citation patterns, then generates better predictions.

-

The use of dynamic topic modeling in our platform allows for better identification of semantic shifts in legal doctrines. The tests confirm that citation networks are characterized by a timeline indicating that evolution in legal language generally occurs before structural evolution. The latter fact transforms the semantic deviation into an early indicator of legal evolution.

-

The accuracy of the obtained causal evaluation confirms that causal reasoning considerably enhances the predictions and their explainability. In legal contexts, this is relevant since the decisions must be justified based on legitimate case law.

-

The use of GNNExplainer allows a better capture of the most important precedent links guiding the prediction process. In our evaluations, this accuracy of interpretation is validated by human experts.

However, despite the good results on the two datasets used, our framework has slightly lower performance on data from the LePaRD dataset, likely due to its focus on semantic retrieval rather than citation structure. Overall, the findings of the study confirm the usefulness of predictive jurisprudence in explaining legal change.

Moreover, despite the robustness and cross-jurisdiction experiments, several limitations remain. First, the controlled perturbations used in the robustness experiments do not capture every possible form of data degradation, including systematic missing citations, changes in legal reporting practices, or sudden changes in doctrinal terminology. Second, the performance of the framework may decrease under a substantial domain or jurisdictional shift, particularly when citation structures and legal conventions differ considerably from those represented in the training data. Third, sparse citation histories and semantically ambiguous precedents remain challenging cases, as reflected by the error analysis in Table 19. Finally, predictions generated by the framework should be regarded as decision-support information rather than substitutes for professional legal judgment. Future deployment-oriented studies should therefore evaluate the framework using continuously updated legal streams, additional jurisdictions, larger distribution-shift scenarios, and prospective human-in-the-loop evaluation.

6. Conclusions

This study proposed a novel AI-based framework for predictive jurisprudence to forecast and explain precedent drift in dynamic legal environments. This framework combines temporal graph neural networks and causal inference with dynamic topic modeling to offer an interpretable solution for analyzing legal evolution. The findings of the study demonstrate that the proposed solution outperformed other algorithms in terms of identifying causal effects, estimating law evolution, and prediction accuracy. On the other hand, advances in artificial intelligence interpretation techniques improve accuracy in legal inference predictions. This progress can be leveraged in numerous applicable fields, such as conducting legal comparisons between countries, utilizing the power of large linguistic models to enhance the semantic understanding of our model, and simulating the outcomes of planned legal reforms.