Libertas: Innocence Asserted

Inspiration

Libertas: Innocence Asserted is a project that seeks to support attorneys at The Innocence Center and beyond in screening transcripts of court proceedings and cases of people who have been convicted of crimes to identify wrongful convictions and advance exonerations. Inspired by the work of The Innocence Center and their success in freeing over 40 people who were wrongfully convicted of crimes, saving collectively over 500 years of prison spent for crimes that they did not commit, this project streamlines the case review and case creation process so that attorneys can focus more time into fighting for justice instead of parsing through hundreds of pages of court transcripts and case documents. It is estimated that between 2% to 10% of all convicted people in the U.S. prison system are wrongfully convicted, suggesting that tens to hundreds of thousands of people are locked away regardless of the truth. This injustice is often buried in legal proceedings and difficult to uncover by organizations and attorneys who already face a mountain of cases to climb in front of them. Often, folks who have been wrongfully convicted of crimes may not know that there are resources available to them that can help exonerate them. Libertas will help attorneys at The Innocence Center and beyond in identifying and reaching out to these people to jumpstart the exoneration process by compiling the information needed to support their case. Libertas: Innocence Asserted streamlines this process based on work done by The Innocence Center, but the model can be expanded to other states and projects focused on freeing those who have been wrongfully convicted.

How we built it

Libertas employs six collaborative AI agents that analyze parole transcripts: (1) Document Intake extracts structured data from PDFs, (2) Innocence Claim Detection identifies direct denials vs. minimization with confidence scores, (3) Evidence Analysis cross-references against 3,400+ exoneration cases from the National Exoneration Registry which were web scraped, (4) Bias Detection analyzes commissioner language patterns and victim presence correlation, (5) Legal Reviewer validates against wrongful conviction indicators, and (6) Synthesis generates auto-populated intake forms and case memos. Libertas generates a memo based on the template used by The Innocence Center which can be used by attorneys to decide if a case could be pursued. Libertas also begins the intake form created by The Innocence Center and completes relevant information based on what is in the transcripts and any other information about the defendant. The attorney dashboard provides a priority queue of ranked cases, searchable transcripts with citations, and bias analytics visualizations, and all outputs include verifiable links to source excerpts.

Challenges we ran into

Throughout the transcripts, sometimes people don't outright say that they are innocent in those exact words. There are so many ways to phrase the "I didn't do it" piece of their story, so it can be difficult to identify the range of language used to describe innocence. Scraping information about exonerations was difficult because of the way the data was formatted on the National Registry of Exonerations. When identifying phrasing or words used to identify victim-focused language or potential bias, it can be difficult to pinpoint every instance of misconduct. Real transcripts also ranged from clean digital text to degraded OCR with garbled speaker labels, still needed some fine-tuning and manual review.

Accomplishments that we're proud of

Using 6 agents worked well so that the modular prompts enabled rapid iteration on specific tasks without breaking others. We are proud of the ability of Libertas to rapidly create memos and fill intake forms to help attorneys identify and speed up cases against wrongful convictions. Combining individual case prioritization with systemic bias pattern detection across thousands of hearings, while generating attorney-ready deliverables to speed case review time will save attorneys time and hopefully lead to more exonerations.

What we learned

We learned throughout this process that transcripts and court proceedings are often kept from public record, locking away the innocence that can be hiding in plain sight when given the chance to be listened to. Attorneys may struggle to review hundreds of pages of transcripts and may miss key patterns or phrases that could indicate innocence, simply because the amount of transcription is too much to process alone. Libertas supports this process by using data and past exoneration examples to identify the innocence that is hidden within these transcripts and compiles this info to then be used by the attorneys to connect with and advocate for those who have been wrongfully convicted.

What's next for Libertas: Innocence Asserted

Libertas: Innocence Asserted can be expanded beyond California by using similar transcript records and can be used by other organizations like The Innocence Center to support their work in pursuing justice. The Open-source models can also be fine-tuned further to reduce API costs, and batch processing can be implemented for 10,000+ historical transcript archives. Audio transcript processing for unwritten hearings could also be added. A public "Parole Transparency Dashboard" with anonymized aggregate bias metrics for policy advocacy could be created to further support exonerations across the country.

Built With

• beautifulsoup4
• gemini
• pypdf2