TL;DR
Researchers at Tulane University have developed an end-to-end data pipeline and dashboard that transforms fragmented, unstructured criminal court records into "analysis-ready" case trajectories. By harmonizing siloed data from New Orleans courts, the system enables community stakeholders to identify systemic patterns in bail inequities, racial disparities, and procedural delays that were previously hidden behind noisy clerk notes.
What Problem Does This Paper Address?
In many U.S. jurisdictions, criminal court data is a "black box." While public portals often exist, they are typically designed for looking up individual cases rather than systemic analysis. In New Orleans, the problem is compounded by a split system: defendants enter through Magistrate Court for initial hearings before being transferred to Criminal District Court. These two systems use different identifiers and do not share a common API.
Furthermore, the most critical data—such as whether a defendant was detained pre-trial or why a case was delayed—is often buried in "docket minutes." These are free-text notes written by court clerks using inconsistent jargon and abbreviations. This fragmentation makes it nearly impossible for journalists, advocates, or researchers to track high-level trends like "charge stacking" (filing multiple overlapping charges to pressure a plea) or racial bias in sentencing.
Methods
The authors built a multi-stage pipeline to turn these "noisy" records into a structured database:
- Data Ingestion & Harmonization: An automated scraper collects raw HTML reports. The system links Magistrate and District court records by cross-referencing arrest charges and searching for embedded magistrate numbers within District Court text entries.
- Rule-Based Event Extraction: Instead of using opaque Large Language Models (LLMs), which can "hallucinate" dates or misinterpret legal boundaries, the team used 21 jurisdiction-tuned regular expressions. These extract key events like "Bond Set," "Charges Refused," or "Guilty Plea Entered."
- State-Graph Modeling: The system maps these extracted labels onto an explicit state graph (see Figure 5). This allows the system to track a case's progression from "First Appearance" to "Case Closed," even when certain procedural steps are skipped in the records.
- Custody Reconstruction: Since pre-trial detention isn't a structured field, the system reconstructs "custody intervals" by scanning clerk notes for linguistic cues like "jail lists," "transport failures," or "released on recognizance."
- Case Similarity: To help users find comparable cases, the system uses a Doc2Vec model. Each case is treated as a document of tokens (charges, demographics, bond amounts, clerk text), allowing the dashboard to surface "nearest neighbor" cases for any given docket.
Datasets
The system was built using public docket records from Orleans Parish, Louisiana. The resulting dataset includes:
- 130,640 cases (125,588 Magistrate and 27,521 District Court cases).
- 820,564 individual proceedings spanning from 1992 to 2026 (based on the paper's temporal scope).
- 2,137 unique charge names.
Key Insights and Ideas
- The "Plea Pressure" of Detention: A striking finding was that 77.4% of defendants held in pre-trial custody eventually pleaded guilty, compared to only 56.8% of those who were not detained. This quantifies the coercive nature of pre-trial detention in driving plea deals.
- Racial Disparities in Delay: Black defendants experienced significantly longer "pending" statuses. On average, their cases lasted 16.1% longer than those of white defendants (90 days vs. 77.5 days).
- The Impact of Charge Stacking: The data validated concerns about "stacking." In single-charge cases, 27.1% of charges were dropped. However, in dual-charge cases, dismissals plummeted to just 9%, while conviction rates rose.
- Explainability over Complexity: The authors intentionally chose deterministic regex over LLMs for event labeling. This ensures that the logic is inspectable—a crucial requirement for a tool intended for public oversight and legal advocacy.
Key Results
The system successfully visualized complex legal trajectories that were previously invisible:
- Gender Disparities: Female defendants were found to be 20% more likely to have all charges dismissed compared to males.
- Bond Outliers: While median bond amounts for battery charges were similar across races, Black defendants faced a much wider range of "exceptionally high outlier bonds," significantly inflating their average bond amount compared to white defendants.
- Dashboard Utility: The final interactive dashboard (see Figure 6) allows users to filter by demographics and charges to see real-time Sankey diagrams of case flows and scatter plots of procedural delays.
Limitations and Future Work
The authors acknowledge several areas for improvement:
- Geographic Scope: The current system is tuned specifically for New Orleans. While the methodology is generalizable, the regex extractors require "jurisdiction-specific tuning" to handle local jargon.
- Data Quality: The system is ultimately at the mercy of the court clerks. If a clerk fails to record an event or uses a completely novel phrasing, the pipeline may miss it.
- Future Integrations: The team plans to integrate police reports and witness data to provide a more holistic view of the "front end" of the criminal legal system. They also aim to refine the Doc2Vec similarity measure to better weight legally significant factors over general text.
Summary updated