Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Designing AI for Prosecutorial Governance: Case Prioritization and Statutory Oversight in Mexico

Fernanda Sobrino, Adolfo De Unánue T., Edgar Hernández, Patricia Villa, Elena Villalobos, David Aké, Stephany Cisneros, Cristian Paul Camacho Osnay, Armando García Neri, Israel Hernández

TL;DR

The paper addresses prosecutorial backlogs in Mexico by developing a dual-function ML decision-support tool for the MAT unit that prioritizes cases likely to close within six months while screening for cases potentially beyond statutory time limits. It uses longitudinal event data from the PIE system and tests multiple classifiers, with Random Forests achieving a mean Precision@300 of about 0.74 under rolling, real-time validation. The system is designed to be low-overhead and non-disruptive, enabling immediate prioritization and governance oversight, and is accompanied by an planned randomized controlled trial. The work also discusses limitations of administrative data, ethical considerations, and the need to adapt the framework to different jurisdictions.

Abstract

Prosecutors across Mexico face growing backlogs due to high caseloads and limited institutional capacity. This paper presents a machine learning (ML) system co-developed with the Zacatecas State Prosecutor's Office to support internal case triage. Focusing on the Módulo de Atención Temprana (MAT) -- the unit responsible for intake and early-stage case resolution -- we train classification models on administrative data from the state's digital case management system (PIE) to predict which open cases are likely to finalize within six months. The model generates weekly ranked lists of 300 cases to assist prosecutors in identifying actionable files. Using historical data from 2014 to 2024, we evaluate model performance under real-time constraints, finding that Random Forest classifiers achieve a mean Precision@300 of 0.74. The system emphasizes interpretability and operational feasibility, and we will test it via a randomized controlled trial. Our results suggest that data-driven prioritization can serve as a low-overhead tool for improving prosecutorial efficiency without disrupting existing workflows.

Designing AI for Prosecutorial Governance: Case Prioritization and Statutory Oversight in Mexico

TL;DR

The paper addresses prosecutorial backlogs in Mexico by developing a dual-function ML decision-support tool for the MAT unit that prioritizes cases likely to close within six months while screening for cases potentially beyond statutory time limits. It uses longitudinal event data from the PIE system and tests multiple classifiers, with Random Forests achieving a mean Precision@300 of about 0.74 under rolling, real-time validation. The system is designed to be low-overhead and non-disruptive, enabling immediate prioritization and governance oversight, and is accompanied by an planned randomized controlled trial. The work also discusses limitations of administrative data, ethical considerations, and the need to adapt the framework to different jurisdictions.

Abstract

Prosecutors across Mexico face growing backlogs due to high caseloads and limited institutional capacity. This paper presents a machine learning (ML) system co-developed with the Zacatecas State Prosecutor's Office to support internal case triage. Focusing on the Módulo de Atención Temprana (MAT) -- the unit responsible for intake and early-stage case resolution -- we train classification models on administrative data from the state's digital case management system (PIE) to predict which open cases are likely to finalize within six months. The model generates weekly ranked lists of 300 cases to assist prosecutors in identifying actionable files. Using historical data from 2014 to 2024, we evaluate model performance under real-time constraints, finding that Random Forest classifiers achieve a mean Precision@300 of 0.74. The system emphasizes interpretability and operational feasibility, and we will test it via a randomized controlled trial. Our results suggest that data-driven prioritization can serve as a low-overhead tool for improving prosecutorial efficiency without disrupting existing workflows.
Paper Structure (16 sections, 2 equations, 8 figures, 5 tables)

This paper contains 16 sections, 2 equations, 8 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Institutional Context: The Zacatecas Prosecutor’s Office
  3. Data
  4. Features
  5. Methodology
  6. Results
  7. Feature Importance
  8. Oversight Function: Identifying Potentially Prescribed Cases
  9. Implementation and Future Steps
  10. Discussion
  11. Ethical Considerations and Limitations
  12. Construction of the Historical Base-Rate Baseline
  13. Sensitivity to Alternative Statutory Prescription Thresholds
  14. Best Model per Algorithm
  15. Best Random Forest Model Groups
  16. ...and 1 more sections

Figures (8)

  • Figure 1: Precision@300 for best model per algorithm across evaluation periods.
  • Figure 2: Out-of-time Precision@300 of the best-performing Random Forest models evaluated on data through January 2025.
  • Figure 3: Aggregated feature importance over time by feature group for Random Forest models.
  • Figure 4: Top 20 most important features in the best-performing Random Forest models.
  • Figure 5: Mean number of potentially prescribed cases among the bottom 1000 cases under the mean statutory threshold, by model family.
  • ...and 3 more figures