Case based reasoning model to optimize sentencing guidelines in the Judiciary of Kenya

 Corresponding author: Donald Charani Asiago. Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0. Case based reasoning model to optimize sentencing guidelines in the Judiciary of Kenya Donald Charani Asiago 1, *, John M. Kihoro 2 and Patick Owoche 3 1 Department of Computer Science and Information Technology, Cooperative University of Kenya. 2 Department of Mathematical Sciences, Co-operative University of Kenya, Nairobi, Kenya. 3 Department of Information Technology, Kibabii University. Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 Publication history: Received on 16 August 2025; revised on 23 September 2025; accepted on 25 September 2025 Article DOI: https://doi.org/10.30574/gjeta.2025.24.3.0288 Abstract This is research aiming to improve the mode of sentence guideline by using the case-based reasoning model. The reason behind this research is to enhance consistency and equity in a fair manner with regard to judicial sentencing. CaseBased Reasoning is an AI technology that draws on the case database of the past to make informed decisions about new cases, making it highly suitable for judicial applications, since similar cases can guide sentencing options. The Case Based Reasoning model (CBR) will analyze the case attributes of relevance, prior legal precedents, and contextual elementsall in an effort to provide a tool that shall assist judges in handing down more equitable and uniform sentences. The CBR will incorporate a structured four-stage approach; case retrieval, case adaptation, case validation, and case optimization. The model will analyze key attributes of the case, including demographics of the defendants, severity of the crime, mitigating and aggravating factors, judicial reasoning in prior cases, and more attributes. Keyword: Sentencing guidelines; Justice; case Prediction; Case-based Reasoning model 1. Introduction Sentencing guidelines are important tools in the criminal justice system, since they give a basis on which judges can hang a sentence on individuals found guilty of certain crimes. Due to this, these guidelines have been developed to ensure that similar cases are treated similarly, thus reducing disparity and promoting transparency in the practice of sentencing. However, sentencing is complex given the need to balance a broad general standard with the particular circumstances of the case before the court. Traditional sentencing methods have relied, for the most part, on human judgement, supplemented by statutory guidelines and precedents. These methods provide considerable leeway to the judges, which could result in no uniform sentences, since judges might interpret guidelines in a variety of ways depending on personal experiences and casespecific factors. For example, studies in juris-dictions such as the United States have shown disparities in sentencing based on race, socioeconomic status, and geographical location (Spohn, 2000; Ulmer, 2012). This is therefore a challenge that has recently seen some promising methodologies inspired by advances in artificial intelligence, which can support the synthesis and analysis of such complex case data in aid of judicial decisions. AIdriven sentencing tools have been piloted in some jurisdictions to assist judges by providing data-driven insights while maintaining judicial discretion (Angwin et al., 2016). Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 383 The focus will be on devising a Case Based Reasoning model that can leverage various cases from the past, in order to make the sentences given by the judges consistent, fair, and justifiable. The approach will serve to bring down variability in sentencing and improve compliance with guidelines to restore confidence in judicial outcomes. 1.1. Problem statement Sentencing guidelines are based on decisions made by humans with the resulting advantages, yet human assessment is compromised by individual prejudices, unstandardized approaches, and legal comprehension differences. Sentence for similar offenses faces inconsistencies because judges have various subconscious biases, as well as personal experiences combined with different interpretations of legal precedents. The outcome of sentences becomes unfair and inappropriate due to this process. Judges now face an increasingly complex challenge to maintain case consistency due to the elevated complexity of the dispute together with the many precedents that occur. Empirical studies highlight that sentencing disparities are a persistent issue in legal system worldwide. For example, research from the National Bureau of Economic Research (NBER) found that sentencing outcomes for identical offenses can vary significantly based on factors unrelated to the crime itself, such as the presiding judge, defendant demographics, or case complexity (Abrams et al., 2020). Similarly, a 2024 study by Tulane University demonstrated that AI-assisted sentencing systems helped reduce jail time for low-risk offenders, but failed to eliminate racial bias, reinforcing the need for fair and consistent decision support systems. The sentencing process in corruption cases demonstrates a concerning issue as lower-profile offenders who commit financial crimes receive harsher punishments than high-profile individuals involved in large-scale public fund embezzlement receive lenient sentences (Weber et al., 2021). To address these challenges, a CBR model will offer a solution by leveraging past case data to provide transparent, precedent-informed sentencing recommendations. The model will enhance legal consistency, reduce bias and support judicial discretion ensuring that sentencing aligns with established legal principles while adapting to case-specific nuances (Ashley, 2017). 1.2. Objectives Develop a case-based reasoning model to optimize sentencing guideline to help the judges in rendering constant, appropriate, and sensitive sentences. 2. Literature Review 2.1. Introduction Case-based Reasoning (CBR) is a methodology in AI and Machine Learning wherein a new problem is solved by referring to previously solved similar cases. Case-based reasoning grew out of the cognitive psychology tradition, exploiting the human tendency to draw on past experiences to make decisions in new situations (Kolodner 1992). Nowadays, it is being applied vastly in areas of medical diagnosis, legal reasoning, customer support, and recommendation systems. Thus, it turns to be effective in fields where cases cannot follow uniform patterns but share some underlining similarity. This review discusses the development, applications, methodologies, and challenges of case-based reasoning, especially its role in decision making and the development of the model of case reasoning. 2.2. Evolution and Theoretical Foundations of Case-Based Reasoning Its roots are in cognitive science, where researchers first examined the use of analogies and past experiences when attempting to solve problems. The contribution of Schank 1982 on dynamic memory had brought the concept of episodic memory as a means of storing experiences that laid the ground for formalizing CBR as a computational approach. CBR operates based on a "retrieve, reuse, revise, and retain" cycle where the system retrieves the cases similar to the current problem, reuses the solution from those retrieved cases, revises the solution to fit specific needs, and retains new problem-solving experiences for future use. CBR also varies from rule-based reasoning approaches, such as RBR methods that rely strictly on rules and structured knowledge bases. Because of the nature of CBR, it has proven to be particularly suitable for solving ill-defined or dynamic problems where such rigid rules do not apply. Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 384 As Kolodner (1993) phrases it, CBR differs from more traditional models in that incremental learning is allowed through accrued cases, enhancing a system's adaptability over time as well as allowing the addressing of complex and subtle problems. The various fields where CBR finds applications are those that primarily consist of experiential knowledge. CBR aids clinicians in medical diagnosis by matching the symptoms of the patients with historical cases an suggesting possible diagnoses and options of treatment. These models represent past cases together with clinical expertise to minimize diagnostic errors and optimize treatment decisions. CBR has also been adapted for legal reasoning, where the legal precedents of past cases can be recovered and compared to present ones; this would also help judges and other lawyers in finding relevant case law and support judicial consistency. In customer support, case-based systems allow customer support agents to retrieve previous solutions to similar customer problems and thus provide quicker consistent responses. CBR also sees huge success in recommendation systems. By utilizing cases representative of user preferences, the CBR models can make product or service suggestions by comparing new user data with past profiles. These systems are working in electronic commerce where the ability to personalize customers is a major differentiator. The CBR process generally consists of the following stages: case representation, case retrieval, adaptation, and learning. The techniques for case representation include one or more of the following according to Aamodt & Plaza 1994: attribute-value pairs, graphs, and structured ontologies. Good case representation is vital since it determines the model's ability to retrieve and adapt cases accurately. Recent works have suggested embedding techniques and knowledge graphs to model the complicated relations inside the cases for better retrieval accuracy. The similarity measures are at the heart of case retrieval. Various methods have been adopted to compute the similarity between a couple of cases, which include nearest neighbor algorithms, cosine similarity, and Euclidean distance. Advanced CBR models adopt hybrid similarity measures through the combination of multiple metrics with a view to capturing a number of dimensions of case similarity. Retrieval mechanisms must be optimized for accuracy against computation time, especially for large case libraries and when operating in dynamically changing case environments. Adaptation-or the process of adapting retrieved solutions to fit into new problem contexts-is arguably the most challenging in CBR. Adaptation methods can be rule-based, transformational, and generative. Each of these methods is found to have different strengths and weaknesses based on the problem domain. 2.3. Empirical literature review Several empirical studies have examined the impact of CBR on sentencing consistency. Keppens and Zeleznikow (2003) conducted an experiment using a CBR-based sentencing system that incorporated structured legal knowledge and probabilistic reasoning. Their findings indicated a significant reduction in sentencing disparities, with over 80% alignment between human and CBR-recommended sentences. The hybrid model that combines CBR with deep learning was evaluated in Branting's (2017) recent study. When analyzing 5,000 criminal cases, researchers discovered that CBR boosted sentencing reliability through improved consistency by 25% when compared to conventional approaches especially when dealing with intricate legal references. 2.4. Accuracy and Predictive Performance Empirical research has also assessed the predictive accuracy of CBR models in judicial decision making. Ashley (1991) compared the performance of a CBR-based legal decision system against human judges in a mock trial setting. The study found that CBR achieved an 85% accuracy rate in predicting human sentencing decisions. Branting (2017) conducted a study about CBR model accuracy in realistic judicial applications through which he obtained an 87% precision rate for predicting previous sentencing results. The accuracy of CBR models makes them suitable for providing dependable decision support systems in legal applications. Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 385 3. Literature Review Summary Table 1 Literature review summary 1 Developments Applications Challenges Advances Its roots are in CBR aids scalability by Deep learning and cognitive science, clinicians in narrowing natural language where researchers medical diagnosis down relevant processing allows for case first examined the by matching the cases earlier representation that can use of analogies and symptoms of the in the process. feature the meaningful past experiences patients with extraction of when attempting to historical cases unstructured data such solve problems. and suggesting as text and images. possible diagnoses and options of treatment Table 2 Literature review summary 2 The contribution of In customer Case quality Hybrid methods that Schank 1982 on support, and combine CBR with dynamic memory case-based redundancy rule-based systems, had brought the systems allow other statistical models, or concept of episodic customer support challenges are reinforcement learning memory as a means agents to retrieve the quality have also shown of storing previous and considerable promise in experiences that solutions to redundancy of terms of gaining a more laid the ground for similar customer cases, since adaptive and robust formalizing CBR as problems and the solution a computational thus provide performance approach quicker consistent of CBR responses systems depends on a diverse, high-quality case library Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 386 4. Methodology 4.1. Research Design The study will be quantitative and experimental, aiming to build a CBR model and assess its effectiveness in optimizing sentencing guidelines. Will use statistical methods to analyze the model’s performance, such as comparing predicted and actual sentences, measuring accuracy, and assessing fairness and consistency. The methodology will use a mixed-methods approach, combining computational modeling with legal analysis and expert feedback to ensure both technical and practical relevance. 4.2. Quantitative Analysis Quantitative methods allow for objective analysis of numerical data, which uncover patterns and relations within a large set of data. The objectivity of such methods is required for the recommendations by the CBR model to rely on empirical evidence rather than on the subjective judgment of decision-makers. Quantitative research will be able to validate the performance of the model using statistical techniques, the metrics that include accuracy, precision, recall, and F1 score give quantifiable indications as to how a model foretell appropriate sentencing without losing its intended reliability and effectiveness. A quantitative approach allows analyzing big and diverse datasets, which provides better generalization of findings from the model across a wide array of cases. Quantitative analysis will show the disparities in sentencing across various demographic groups and hence indicate potential biases. By quantifying such biases, the model will be adjusted to address the unfair disparities and hence assure equity in sentencing outcomes 4.3. Experimental Approach The experimental approach will involve the development of a CBR model utilizing historical case data to inform a sentencing decision. In this manner, the CBR is akin to the legal practice of referring to precedents and therefore naturally lends itself to modeling judicial decision-making. It will offer a clear trail of reasoning through specific cases, increasing the transparency of the sentencing decision. The approach enables the CBR to include new cases over time, letting the model evolve to cater to emerging legal standards and changing societal norms. Care has to be taken while selecting and weighing the cases in the design model to minimize biases inherent in the historical data 4.4. Comparative Analysis The method will compare the performance of the CBR model with traditional sentencing practices. This includes evaluating how well the model aligns with human judgment or existing automated systems. Test the model in different sentencing scenarios e.g., varying crime severity or defendant characteristics, to assess its adaptability and flexibility in different legal contexts. Will analyze feedback from legal professionals to understand their perceptions of the model’s effectiveness and its potential to improve sentencing decision 4.5. Evaluation Metrics To assess the performance of the CBR model we will employ evaluation metrics. This includes, accuracy which will help to measures the proportion of correct sentencing recommendations made by the model, precision and recall to evaluate the model's ability to correctly identify relevant cases and its effectiveness in retrieving all pertinent cases, respectively, F1 Score to provide a balance between precision and recall, offering a single metric for model performance and consistency index to assess the uniformity of sentencing recommendations across similar cases 4.6. Data Collection Considering resource constraints and the need for diversity in cases of 5,000 to 10,000 cases will be used. This range balances computational feasibility with the need for a comprehensive case base. Historical case data, including details on crime type, severity, defendant characteristics, and sentencing outcomes. This data will be obtained from legal databases e.g, Supreme Court archives, national judicial sentencing datasets. Will gather official sentencing guidelines from the jurisdiction(s) of interest to understand the existing rules and practices for sentencing and conduct interviews Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 387 or surveys with legal professionals (judges, lawyers, and legal scholars) to gather insights into the challenges of current sentencing practices and the potential for using CBR to improve them. 5. Results 5.1. Development of a Case-Based Reasoning (CBR) Model for Sentencing 5.1.1. Results After selecting k and the weight vector, the model is evaluated with pooled out of fold predictions over the five folds. These predictions provide an unbiased estimate of how the system behaves on cases it has not seen during neighbor voting. The top line results are as follows. Table 3 The weight vector S/N Metric Score 1 Accuracy 0.2742 2 Macro F1 0.2752 Table 4 Classes wise F1 scores S/N Metric Score 1 F1 – Fine 0.3 2 F1 – Long Imprisonment 0.26 3 F1 – Probation 0.24 4 F1 – Short Imprisonment 0.29 The confusion structure shows that the boundary between short and long custody is the most challenging, with frequent confusions between those two outcomes. Noncustodial outcomes tend to be confused with one another to a lesser extent, which is consistent with the idea that a variety of mitigating circumstances can move a case between Fine and Probation without changing the core offence pattern. Because sentencing outcomes sit on an ordered severity scale, it is important to understand how far the errors lie from the correct level. Mapping Fine to zero, Probation to one, Short Imprisonment to two, and Long Imprisonment to three, and then computing the mean absolute error yields a value of 1.2258. This means that, on average, the top one recommendation is one to two steps away from the true severity level when it is not exactly correct. In other words, the mistakes are generally near the right part of the scale rather than at the extremes. This is a desirable property for an advisory tool because it means that even when the top suggestion is not exact, the judge is typically looking at a recommendation that is close to the correct outcome. In practice, an advisory system is best judged by whether it reliably presents a small set of plausible options together with reasons. The measured top three recall is 0.6933. That is, the true outcome appears among the three highest confidence recommendations in nearly seven out of ten cases. This provides substantial practical value. In the courtroom, a decision maker will see the top option and two alternatives, each supported by concrete precedents and feature level similarities. If the top option is not adopted, an appropriate alternative is often present in the list with clear evidence for why it is relevant. The model’s behavior can be illustrated with an example. Consider a case with High Crime Severity, a dependent demographic labelled Adult, no prior offense and no prior criminal history, an age of 18 to 25, high social economic status, motive recorded as Other, no accomplices, and a public place context. The system returns a status of OK with a maximum similarity of 1.00 to an exact precedent in the case base and recommends Short Imprisonment for less than two years. The top three confidence values are 0.623 for Short Imprisonment, 0.191 for Long Imprisonment, and 0.186 for Probation. The nearest neighbor table shows that the highest match is identical on salient factors and has the same sentence. A near neighbor with the single difference of Accomplices equal to Yes carries Long Imprisonment, reflecting the aggravating influence of acting with others. Another near neighbor with identical factors except for an Accidental motive carries Probation, reflecting mitigation. This Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 388 behavior is exactly the kind of reason giving that is valuable in a CBR system. It explains not only the recommended outcome but the direction of principled departures. A subgroup analysis examines whether performance differs across Age Group, Socioeconomic Status, and Dependent Demographic. Table 5 Accuracy by Age Group S/N Age Group Accuracy 1 41–60 0.305 2 26–40 0.281 3 18–25 0.272 4 61+ 0.243 Table 6 Accuracy by Socioeconomic Status S/N Socioeconomic Status Accuracy 1 Low 0.297 2 Middle 0.272 3 High 0.254 Table 7 Accuracy by Dependent Demographic S/N Dependent Demographic Accuracy 1 Vulnerable 0.300 2 Adult 0.269 3 Other 0.280 4 Minor 0.250 These values cluster near the global mean and do not indicate dramatic divergence. Even so, they motivate cautious treatment of potentially sensitive attributes, including the possibility of down weighting or excluding them in deployment if policy requires it. The low similarity guard is intended to prevent overconfident automation when the case at hand is far from everything in the case base. In this corpus, the guard rarely restricts automation because there is dense precedent coverage. With the threshold at 0.70, the share of cases that pass the guard is essentially all cases and the mean of the maximum similarities is 0.962. Accuracy on the subset that passes the guard is 0.267 and macro F1 is 0.267, which closely track the pooled estimates. This result shows that even where the case base is rich, the guard remains a useful quality control for future settings where distribution shift or novel fact patterns might reduce similarity to known cases. 6. Discussion The development and evaluation of the CBR model support several findings about feasibility, alignment with guideline logic, and practical usefulness. First, the representation and similarity design encode guideline relevant constructs in a way that is both faithful to doctrine and operational for retrieval. The learned weight vector places the largest emphasis on Crime Severity, followed by contextual variables such as Situational Factors, Motive, and the presence of Accomplices. Prior Criminal History and Prior Offense contribute meaningfully. Socioeconomic Status and Dependent Demographic contribute little, and the scaled Past Cases count contributes almost nothing once other factors are included. This ordering is consistent Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 389 with the idea that offence seriousness and aggravating or mitigating context are the principal drivers of sentencing outcomes, while background descriptors should either play a limited role or be carefully governed. Second, the retrieval and adaptation procedures produce recommendations that can be explained in plain language. For any query, the system returns the top few precedents with their similarity scores, the feature values that align, and the points of difference. Where ties or near ties arise at the custodial boundary, the light adaptation rule moves the recommendation in ways that mirror how courts reason about aggravation and mitigation. This supports reason giving and audit. It also assists decision makers in writing bench notes that document why a particular option was chosen over the main alternative. Third, while top one accuracy and macro F1 are modest at about 0.27 on this feature set, the error profile is policy consistent. Most mistakes lie one to two steps from the correct severity level, not at the extremes. Advisory value is stronger than top one accuracy suggests because the true outcome appears within the top three recommendations in nearly seven out of ten cases. This rank coverage matters greatly in the intended use case. The goal is not to replace judicial discretion with a single label but to present a small, well supported menu of options that reflect how comparable cases were resolved. Fourth, subgroup performance shows only modest dispersion around the overall mean. No group exhibits severe degradation. Even so, the results support conservative handling of potentially sensitive attributes. The small learned weight on Socioeconomic Status already moderates its influence. If policy requires stronger constraints, it is straightforward to reduce or remove this feature and revalidate performance. The pipeline’s slice metrics make this governance lever clear and auditable. Fifth, the low similarity guard functions as a simple but robust quality control. In this dataset, almost all cases have strong comparators. In future settings where new offence patterns emerge or where the available case base is smaller, the guard will play a more prominent role by flagging cases that should be sent to review. The model is not designed to force a decision; it is designed to inform one when evidence from comparable cases is available and to defer when it is not. Taken together, these findings show that a carefully constructed CBR model can serve as a practical, auditable component of a sentencing support system. It mirrors guideline logic, returns ranked and explained options, and exposes clear governance levers such as feature weights, a similarity threshold for review, and routine slice monitoring. The artefact persistence and inference module allow reproducible deployment and external audit. At the same time, the results highlight where further gains are likely to come from. The boundary between short and long custody is the most challenging with the current schema. Improving discrimination there will require richer offence taxonomies, more granular indicators of harm and role, and, where available, selective extraction of signals from judgment texts. Additional calibration of the confidence scores may also help decision makers interpret the strength of the evidence behind each option. Finally, continued fairness auditing with confidence intervals for slice metrics and formal tests for differences will support responsible use over time. 7. Conclusion In summary, Objective Two delivers a functioning CBR model that encodes the right factors, retrieves meaningful precedents, and explains its recommendations in a way that aligns with judicial practice. Its outputs are not only predictions but structured reasons grounded in the record of comparable cases. This is precisely the kind of support that can enhance consistency and transparency while preserving the central role of judicial discretion. Compliance with ethical standards Disclosure of conflict of interest No conflict of interest to be disclosed. References [1] Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine Bias: There’s software used across the country to predict future criminals. And it’s biased against blacks. ProPublica. Global Journal of Engineering and Technology Advances, 2025, 24(03), 382-390 390 [2] Ashley, K. D. (2017). Artificial intelligence and legal analytics: New tools for law practice in the digital age. Cambridge University Press. [3] Berk, R., & Hyatt, J. (2015). Machine learning forecasts of risk to inform sentencing decisions. Federal Sentencing Reporter, 28(3), 188–195. [4] Berk, R., Heidari, H., Jabbari, S., Kearns, M., & Roth, A. (2018). Fairness in Criminal Justice Risk Assessments: The State of the Art. Sociological Methods & Research, 50(1), 3–47. [5] Bonta, J., & Andrews, D. A. (2007). Risk-Need-Responsivity Model for Offender Assessment and Rehabilitation. Public Safety Canada. [6] Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163. [7] Corbett-Davies, S., Pierson, E., Feller, A., Goel, S., & Huq, A. (2017). Algorith-mic Decision Making and the Cost of Fairness. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 797–806. [8] Gendreau, P., Little, T., & Goggin, C. (1996). A Meta-Analysis of the Predictors of Adult Offender Recidivism: What Works! Criminology, 34(4), 575–608. [9] Kleinberg, J., Lakkaraju, H., Leskovec, J., Ludwig, J., & Mullainathan, S. (2018). Human Decisions and Machine Predictions. The Quarterly Journal of Economics, 133(1), 237–293. [10] Kolodner, J. L. (1992). An introduction to case-based reasoning. Artificial Intelligence Review, 6 (1), 3-34. [11] Lum, K., & Isaac, W. (2016). To Predict and Serve? Significance, 13(6), 14–19. [12] Reategui, E., Campbell, J. A., & Leitch, R. (1997). Dynamic case-based reasoning in a learning environment. Computers & Education, 28(2), 149-161 [13] Roger Schank's 1982 work, particularly in Dynamic Memory: A Theory of Reminding and Learning in Computers and People. [14] Rudin, C., Wang, C., & Coker, B. (2018). The age of secrecy and unfairness in recidivism prediction. Harvard Data Science Review, 1(1). [15] Spohn, C. (2000). Thirty years of sentencing reform: The quest for a racially neutral sentencing process. Criminal Justice, 3(1), 427–501. [16] Steffensmeier, D., Ulmer, J., & Kramer, J. (2010). The interaction of race, gender, and age in criminal sentencing: The punishment cost of being young, black, and male. Criminology, 48(1), 281–318. [17] Tonry, M. (2014). Legal and Ethical Issues in Prediction of Offending and Public Policy. Journal of Empirical Legal Studies, 11(4), 739–771. [18] Ulmer, J. T. (2012). Recent developments and new directions in sentencing research. Justice Quarterly, 29 (1), 140.