2.1. Jurimetrics and AI in courts

The term jurimetrics, the quantitative analysis of legal processes, was first introduced by Lee Loevinger in 1949 (Loevinger, Reference Loevinger1949, pp. 455–472). It refers to the application of statistics, logic, and computational tools to legal reasoning. Over time, jurimetrics has evolved into computational jurisprudence, leveraging statistical models and natural language processing (NLP) to extract patterns from large legal corpora. Katz et al. (Reference Katz, Bommarito and Blackman2017) developed a predictive model for the US Supreme Court decisions with high accuracy, demonstrating that historical case data can effectively forecast judicial behaviour. Similarly, Chen and Li (Reference Chen and Li2020) and Bench-Capon and Sartor (Reference Bench-Capon and Sartor2003) show that machine learning can analyse judicial behaviour and support automated reasoning. However, most jurimetrics studies target common-law systems. In contrast, civil-law jurisdictions like Indonesia differ substantially in document structure, reasoning style, and institutional transparency. Judicial decisions often follow narrative rather than issue-based formats, posing challenges for computational modelling. Consequently, research on civil-law jurimetrics remains limited, which motivates us to research the development of AI systems tailored to Indonesia’s legal framework.

2.2. International legal analytics

In practice, law firms and governments use AI for litigation analytics. LexisNexis’s Lex Machina and Thomson Reuters’ Westlaw Edge mine millions of legal documents to advise on strategy and predict outcomes (LexisNexis, 2018). Fortunately, their efforts have shed light on judges’ tendencies and likely rulings. These tools exemplify how legal analytics can make research evidence-based and efficient. In China, the Supreme People’s Court has championed a “Smart Court” model by introducing its online judgment database (“China Judgments Online”), which publishes nearly all court decisions (Stern et al., Reference Stern, Liebman, Roberts and Wang2021). As documented by Chen and Li (Reference Chen and Li2020), Chinese courts employ predictive analytics to promote consistency and detect judicial decisions that deviate from established statistical norms. AI-powered tools, such as the “Wise Judge” system, search this database to ensure that similar cases receive similar verdicts. Anecdotally, Chinese courts report faster dispositions. For example, Internet Courts, such as those in Hangzhou, have reported case resolution times reduced by nearly half compared to traditional trials. These developments illustrate uniformity of justice and greater transparency, which public case data allows for scrutiny. The quoted vision is that AI “will reduce judicial arbitrariness and promote justice” by guiding judges towards objective evidence-based decisions.

Nonetheless, experts caution that AI in courts must be carefully managed due to its bias detection and explainability. For instance, some Chinese studies note that overreliance on AI could discourage judges from fully articulating their reasoning (Zhang, as cited in Shi, Reference Shi2021). These international experiences provide important reference points, but cannot be directly transplanted into jurisdictions such as Indonesia, as legal doctrine, institutional design, and expectations of judicial discretion differ substantially. In summary, global developments demonstrate AI’s potential for enhancing judicial efficiency and consistency, while also highlighting the need for local adaptation, oversight, and respect for judicial norms.

2.3. Text extraction in law

From a technical standpoint, legal text mining combines rule-based methods, statistical models, and increasingly large neural networks. Rule-based extraction using regular expressions and dictionaries is precise for structured elements, such as case numbers, dates, and citations, but brittle to unstructured long text summarisation. Transformers such as BERT and domain-adapted variants such as Legal-BERT provide robust NER and classification, which are good for tasks such as citation extraction and issue tagging. However, most benchmarks are on English corpora (Chalkidis et al., Reference Chalkidis, Fergadiotis, Malakasiotis, Aletras, Androutsopoulos, Cohn, He and Liu2020). Large-language models offer generative capabilities, enabling open-ended summation and semantic queries.

Practical use of text extraction in legal contexts requires carefully designed prompts, rigorous validation to ensure accuracy, and substantial computational resources when deployed locally. The state of the art suggests hybrid pipelines that combine the high precision of rules with the flexibility of ML/LLM models, under an auditing framework (Francesconi et al., Reference Francesconi, Montemagni, Peters and Pasini2010). This research follows that approach, adapting hybrid extraction strategies to Bahasa Indonesia and Indonesian tax law, where linguistic and structural nuances differ from English-language corpora.

2.4. Research gaps and contributions

The existing literature highlights two major gaps. First, there is no comprehensive extraction tool specifically designed for Indonesian Tax Court decisions, nor any baseline metrics for such data. Second, previous jurimetrics efforts have not effectively integrated Indonesian legal specifics (language and structure) with modern AI techniques. Meanwhile, our ADIL framework directly addresses these deficiencies. It introduces an Indonesian-specific pipeline and establishes baseline performance metrics for extraction and prediction. ADIL operationalises jurimetrics for a civil-law jurisdiction and contributes empirical evidence to the emerging field of Indonesian legal analytics.

3.1. Data collection and preprocessing

All publicly available Indonesian Tax Court decisions from 2008 to 2023 were collected from the public repository maintained by DDTC (https://ddtc.web.id) in HTML format. The raw corpus consisted of 6265 files spanning multiple tax categories: customs duty, income tax, land and building tax, value-added tax, and others, with a total disputed amount exceeding Rp 106.6 trillion. To ensure representativeness and analytical efficiency, we applied the Slovin formula to determine an appropriate sample size from a total population of ~210,000 tax court decisions from 2008 to 2023. Using a 5% margin of error, the calculated sample size was n = 399 cases. Thus, our corpus of 6265 decisions far exceeds this threshold and provides sufficient coverage. The sample was proportionally stratified by year and dispute type (appeal, lawsuit, and judicial review) to preserve heterogeneity across chambers and case categories. The Slovin method was chosen because it is suitable for heterogeneous populations with large datasets, ensuring that the results can be generalised to all tax court decisions issued between 2008 and 2023 (Slovin, Reference Slovin1960; Tejada and Punzalan, Reference Tejada and Punzalan2012).

After quality filtering, by removing incomplete metadata or garbled files, the final dataset comprised 4377 decisions. These judgements vary in length and content structure, reflecting different chambers and years. We split the data chronologically. The latest 20% was held out as a test set for evaluation, with the remainder used for development and cross-validation. All documents underwent extensive preprocessing to standardise textual structure and maintain auditability. We first cleaned the HTML by removing irrelevant tags, scripts, and styling, and normalised text (Unicode standardisation and Indonesian diacritics correction). We then implemented document structure detection, using custom rules to identify sections such as procedural history, factual chronology, legal reasoning, and pronouncement. Each preprocessing step was logged to preserve provenance and reproducibility. This multi-stage pipeline ensured that every document entered the analytical phase in a consistent and traceable format suitable for large-scale extraction and machine learning.

3.2. ADIL metadata extraction pipeline

We structured the cleaned text into a fixed schema of 49 metadata fields. These include identifiers (case number and URL), categorical fields (tax object class/subclass and litigation class), temporal fields (assessment year, objection year, decision year, and tax year), financial data (disputed amount and currency), and textual summaries. Critically, we also extract entities, such as the names of judges on the panel, the tax office objection reference, and the Directorate General of Taxes (DGT) appeal number, among others. The targets of interest for analysis are the final litigation outcomes from the government’s perspective: win, lose, or partial win. We put the cases’ duration in years, measured from the objection decree to the court decision.

To perform extraction processing steps, ADIL integrates three complementary approaches, evaluated in a unified framework:

• • Rule-based patterns: We encoded regular expressions and lookup tables for rigid formats. For instance, case numbers, legal citations, and numeric values are highly regular. Rule-based extraction achieved a precision of >94% on structured items.

• • Transformer-based NER: We fine-tuned pretrained multilingual and Indonesian BERT models, IndoBERT, XLM-RoBERTa, and mBERT, on Indonesian legal text to recognise entities, such as person names, organisations, and legal references. These models excel at recall and disambiguating context, complementing the rules. In our experiment, the Transformer-based NER model demonstrated the lowest performance among the evaluated methods.

• • LLM semantic analysis: We also leverage a local large-language model, Gemma3, using carefully designed prompts to summarise case facts and classify dispute taxonomies, such as determining tax type from narrative. The LLM reads each decision and outputs JSON fields like “Tax Type” and “Summary of Facts.” This method effectively handles complex or unstructured information, addressing the biases and limitations inherent in rule-based and NER approaches.

For each field, we determine the extraction output using a simple voting rule. If the rule-based and NER systems succeed in extracting metadata with the correct data types and formats, we accept it. Otherwise, we fall back to the LLM’s answer or flag it for manual review. We evaluated extraction accuracy on a gold-standard subset of 800 manually annotated cases. Table 1 (see the Supplementary Appendix S1) reports per-field precision, recall, and F1 for each method and their combination. The full multimodal pipeline (rule-based pattern + NER + LLM) achieved a precision of 92.3%, a recall of 89.7%, and an F1 score of 91.0% overall. Some key examples include precision for document identifiers and dates, which exceeded 95%, while precision for amounts and statutory citations was 85–90%. Crucially, throughput remains practical for approximately two documents/sec on our hardware, and it allows the entire corpus to be processed in hours.

All extracted data are stored in a structured database with provenance logs. Sensitive data, such as judge names, is anonymised by an asterisk mask in outputs to comply with privacy norms, retaining consistent pseudonyms for analysis. We derived additional features from the raw metadata. For example, the tax assessment letter follows the format AAAAA/BBB/CC/DDD/EE, from which we parsed out: sub-class code (BBB = tax/sub-tax category), tax year (CC), and letter year (EE). We encoded all categorical fields, such as tax codes, case categories, and currency, with integer labels. Continuous variables like disputed amounts were binned into ordinal classes, using Sturges’ rule to reduce skew. We also computed the durations in years from the objection decree to the court decision and validated them against known procedural constraints in the Indonesian Tax Law. Outliers were handled with Tukey’s rule and then imputed by class means. By the end of preprocessing, we had a feature matrix of 4377 cases × 12 top-selected features, ready for modelling.

3.3. Predictive modelling

A primary methodological issue in predicting judicial outcomes is the availability of model inputs before or only after the adjudication process. Medvedeva et al. (Reference Medvedeva, Wieling and Vols2023) point out that many predictive systems depend on post-decision variables, like judicial reasoning or citation patterns. This makes them less useful for real forecasting and decision support.

To counter this criticism, all predictive models in this study are trained and assessed solely on ex-ante variables that are discernible before the issuance of a judicial decision. These are case metadata (type of dispute, tax category, range of disputed amounts, and year of filing), procedural attributes, and the makeup of the panel. The forecasting feature set does not include variables taken from the reasoning text, verdict section, or statutory citations of the final judgement. Instead, these variables are only used for retrospective evaluation, interpretability analysis, and model auditing. In this phase, we build predictive models on the structured data. We consider two supervised tasks:

1. (A) Outcome prediction: a classification task to predict the Government’s case result (“win” vs. “lose” vs. “partial-win”) from pre-decision metadata; and

2. (B) Duration prediction: a regression task to forecast case resolution time (in years). Each task was trained using fivefold cross-validation on the training set and evaluated on the held-out test set.

We tested a range of algorithms, including Gradient Boosting, XGBoost, Random Forest, Neural Network, Support Vector Machines (SVM), and Logistic Regression model for the classification task, and Gradient Boosting, XGBoost, Random Forest, Support Vector Regressions (SVR), Linear Regression, and Ridge Regression. Model hyperparameters were tuned via grid search.

To investigate the influence of judges on tax court decisions, we ran each model twice. Once with the encoded judge name included, and once without the judge name, with all other features held constant. This method isolates the impact of the judge’s factor. We report metrics as follows:

1. (1) For classification, we report accuracy, F1-score, precision, and Recall;

2. (2) For regression, R ² and RMSE. We also record cross-validation means and standard deviations to assess stability.

3. (3) Finally, we examine feature importance using chi-square scores and tree-based importances to identify the most predictive variables.

Figure 1 illustrates the ADIL system architecture, which we implement as a layered pipeline. Each layer has a distinct role as follows:

1. (1) Data layer: Aggregates and prepares legal data.

Figure 1.

ADIL system architecture.

Figure 1. Long description

At the far left, the Data Layer contains two vertically stacked boxes: Data Ingestion with Extraction from Court Decisions and JSON Standardization, and Data Cleaning with HTML or XML Tag Removal, Text Normalization, and Anonymization. Arrows lead rightward to the Analysis Layer, which contains three vertically arranged modules: Jurimetrics Core with Citation Network Analysis and Statistical Modeling, A I Core Gemma3 with Transformer-based N L P and Prompt Engineering, and Temporal Validation with Rolling Window. An arrow leads rightward to the Ethical Compliance Layer, which contains Bias Mitigation, Fairness Assessment, and Bias Monitoring. The final arrow points to the Interface Layer, which contains Decision Support Dashboard, Explainable A I, S H A P, Precedent Visualization, and Outcome Prediction.

We ingest tax court decision documents, statutes, and related precedents and perform rigorous preprocessing. The preprocessing stage includes text cleaning, character normalisation to handle Bahasa Indonesia characters, section segmentation, and language detection. We enforce temporal integrity through a ChronosLex scheme so that models only learn from past data, reducing look-ahead bias. Crucially, sensitive information is anonymised by masking the judges’ names using a context-aware entity to comply with the Indonesian PDPL. To address privacy trade-offs, we quantify re-identification risk by analysing pseudonymised datasets under various scenarios, ensuring compliance with current standards. In case future regulations tighten disclosure rules, we have a contingency plan to adjust pseudonymisation methods or restrict specific data outputs. By producing a clean, chronologically ordered dataset with privacy safeguards, the Data Layer ensures a solid, compliant foundation for analysis.

1. (2) Analysis layer: Performs Jurimetric and AI computations.

We used ChatGPT to produce the equation as a picture that can be seen in Figure 2. SHAP Formula. Shapley Additive Explanations (SHAP) functions as the interpretability engine that unifies statistical reasoning, jurimetric analysis, and AI inference. SHAP originates from cooperative game theory, where each feature in a model is treated as a “player” contributing to the final prediction. The SHAP value for a feature represents its marginal contribution to the model’s prediction. How much that feature increases or decreases the model’s output when considered across all possible combinations of features (Lundberg and Lee, Reference Lundberg and Lee2017). Within the analysis layer, SHAP functions with the following formula:

Figure 2.

SHAP Formula.

This layer integrates quantitative legal analytics with machine learning. For example, we apply citation-network analysis and econometric models of judicial behaviour alongside our Gradient Boosting outcome predictors. Citation-network analysis is operationalised by constructing rolling graphs of statutory and jurisprudential references across time windows. Distribution drift is flagged when statistically significant changes occur in network structure, including shifts in the centrality of frequently cited provisions, the emergence of new citation clusters, or the declining relevance of previously dominant precedents.

In parallel, econometric models of judicial behaviour—implemented using panel data and fixed-effects regressions—track changes in outcome probabilities across judges, panels, and time. Structural breaks in coefficients or unexplained residual shifts serve as indicators of behavioural or institutional drift. When these analytical signals exceed predefined thresholds, the system triggers controlled retraining on recent data.

1. (3) Ethical compliance layer: Oversees fairness and governance.

Sitting above the pipeline, this layer enforces legal and ethical standards at runtime. It logs all processing steps and model decisions for auditability. It continuously monitors algorithmic bias by inspecting SHAP summaries across cases. In this context, a feature group is considered to unfairly skew outcomes if its SHAP attribution demonstrates a disproportionate influence on predictions that cannot be justified by legally relevant case characteristics, and if outcome disparities persist after controlling for substantive legal factors. Legally permissible variables are not suppressed, but such patterns are flagged for further review. It also supports enforcement of data governance rules and periodic algorithmic audits by verifying compliance with Indonesian regulations, such as the Indonesian Personal Data Protection Law, and core judicial ethical principles, including impartiality, equality before the law, and the prohibition of extraneous influence. Compliance is operationalised through audit logs, access controls, and periodic algorithmic reviews. When ethical thresholds are breached, the system mandates human review rather than automated intervention.

1. (4) Interface layer: Presents results to users.

This layer provides a user-friendly dashboard. It visualises the AI’s suggestions alongside human-readable explanations. For instance, when the model predicts a “win/lose/partial win” outcome, the interface displays the top SHAP features for that case and highlights similar precedent cases in a citation graph. We include confidence scores and enable users to query the reasoning behind the system’s recommendations. The design ensures transparency, allowing judges to drill down into the decision path or legal citations underlying any prediction. Even though human–AI interaction is central, the interface never overrides a judge’s authority but serves as a co-pilot to augment judicial reasoning.

Collectively, these layers create an end-to-end decision support framework raw tax litigation data flows through secure ingestion, intelligent analysis, and a transparent interface stage, all under the oversight of an alignment layer that ensures the system adheres to legal norms. This layered design (as depicted in ADIL’s architecture) ensures that the AI tools remain explainable, auditable, and compliant with Indonesian law.

All predictive models in this study rely exclusively on variables that are observable before the issuance of a judicial decision. These include case metadata (dispute type, tax category, disputed amount, and filing year), procedural attributes, and historical jurisprudential indicators available at the time a case is lodged or assigned. Variables derived from the reasoning text or verdict section of the final judgment are explicitly excluded from the forecasting feature set and are used only for retrospective evaluation, performance auditing, and model retraining. This design ensures that the models support genuine ex-ante forecasting rather than post-hoc rationalisation, in line with best practices in judicial prediction research.

3.4. System implementation and hardware specifications

Our ADIL pipeline is implemented in a containerised framework, facilitating seamless deployment. All components run on an adequate workstation equipped with an Intel Core Ultra 9 processor, an NVIDIA GeForce RTX 5090 GPU for acceleration, and 128 GB of DDR5 memory. This hardware configuration was chosen to meet the intensive computational demands of legal NLP and model training. The GPU enables the rapid processing of documents and the prompt-based generation of summaries (for semantic summarisation). In our tests, the system achieved throughput of several documents per second in batch mode. The 128 GB RAM ensures that transformer-NER, local Large Language Model, and large datasets can be accommodated in memory during the training, validation, and prediction processes. We run the framework in Docker containers, facilitating reproducibility and eventual integration into institutional environments.

3.5. Deployment prospects and challenges

The ADIL decision-support system holds several practical benefits for Indonesia’s tax judiciary. Automating metadata extraction and analysis can reduce case backlogs and processing times. Prior studies have noted that AI tools for case review can help judges handle higher caseloads more efficiently (Kanyi, Reference Kanyi2019; Zakaria et al., Reference Zakaria, Ahmad, Hussin, Hassan, Marzuki, Syukur and Sari2024). In our framework, tax officers can quickly retrieve precedent cases, statistical patterns, and outcome predictions at each stage of the case, which potentially leads to more consistent rulings. Taxpayers could proactively identify and mitigate potential risks before initiating any legal action or appeal. For tax consultants, the system enables risk mitigation in managing personnel, time, and financial resources throughout case resolution. The system’s explainability, achieved through SHAP and summaries, also helps build judges’ trust, as decisions are not black boxes but are accompanied by highlighted laws and similar judgements. For the government, this framework facilitates evidence-based evaluation of prevailing tax laws, regulations, and policies, fostering procedural fairness and continuous improvement to minimise future tax litigation.

The deployment stage may face some significant challenges. First, AI models are trained on historical data. If past Tax Court decisions exhibit any unconscious bias, the AI will codify and amplify this bias, undermining the goal of fairness and impartiality. Judges may distrust the AI’s recommendations if the underlying logic is non-transparent, as in the “black box” problem, making it difficult to reconcile the objective data with their subjective conviction. Second, deploying a data-driven system requires cultural change. Judges and administrators may resist adopting technology that appears to encroach on their traditional authority or judgement, especially in a system where subjectivity has been a key weakness. Without adequate training and clear communication, the system may be underutilised or, conversely, over-relied upon, where judges blindly follow AI predictions instead of exercising their essential judicial role. Third, AI requires large volumes of clean, consistent, and well-structured historical data. The complexity and volume of tax dispute cases are high, but the existing digital records may not be standardised or consistently tagged. Poor data quality will lead to inaccurate AI predictions, eroding confidence in the Jurimetrics system, and potentially leading to incorrect case triage.