CommentaryIn this five-part series, data analyst David Stasior argues that AI has made it possible to turn court and disciplinary records into data at scale, allowing the rule of law to be measured rather than just assumed. Part 1 made the case, and Part 2 put it to work on 24 years of Massachusetts attorney disciplinary decisions. Here, in Part 3, he turns to federal court statistics, where AI can quickly uncover insights linked to the rule of law, despite awkward formatting that makes the numbers hard to use. Part 4 points AI at the text of federal court decisions to reveal the reasoning judges use when deciding to end supervised release. In Part 5, the final installment, Stasior turns from hindsight to real time.
Throughout this series, AI has tested our beliefs about the rule of law against empirical data—uncovering consumer-complaint patterns in thousands of attorney disciplinary decisions and testing the federal “rocket docket” reputation.
Yet systemic statistics reveal how courts differ, not necessarily why. In 2025, the Eastern District of Virginia resolved 16 civil cases by trial with a median of 17.9 months from filing, while the District of Utah resolved 10, with a median of 67.7 months. Differences that large raise serious questions about whether litigants are harmed by procedures that move too fast or too slow. To understand what actually drives these disparities, we must look at the case records themselves.
Unfortunately, the paywall of PACER, the federal judiciary’s fee-based online records system, keeps most records essentially hidden. Some documents do reach the public domain, and these decisions provide at least a window into judicial decision-making. Part 3 ended with one real, illustrative case: a request to terminate supervised release early.
On September 30, 2025, the federal system supervised 107,565 people serving these terms, with a median length of about three years. These individuals are out of prison but remain subject to strict conditions governing where they can live, who they can associate with, and how they work. At any time, a probation officer can visit their place of work, search their home, or demand drug testing. Because caseloads are heavy, the probation system is designed to focus resources on those who continue to struggle with the rules, leaving many well-adjusted individuals requiring only minimal, low-level supervision.
Many defendants eventually petition for early termination. Decisions about ending supervision early do not just affect those 107,565 people; they ripple out to families, employers, and probation officers. While some individuals genuinely need ongoing monitoring, a consistent standard for identifying those who do not could safely and significantly reduce the federal supervision footprint.
The statutory standard governing this process, however, is surprisingly vague. After a defendant serves one year of supervised release, federal law permits a judge to terminate the remaining term if the defendant’s conduct warrants it and the decision is in the “interest of justice.” While a judge must also weigh standard sentencing factors — such as the nature of the offense and public safety—the ultimate outcome hinges on that highly subjective phrase.
That is the entirety of the statutory guidance. So how does this play out in the courtroom? If a defendant finds work, stays sober, and complies with every rule, does that justify early termination? Or is that simply what supervision requires in the first place?
Answering that system-wide is nearly impossible because PACER’s paywall hides thousands of unpublished orders. But we can use the fraction of decisions in the public domain to test how judges apply the law. Because AI is uniquely suited to extract the specific reasoning buried inside unstructured legal text, we can use it to map the logic driving these outcomes. Running on an ordinary Mac mini, an AI agent processed these individual decisions and isolated their governing logic in about five seconds each.
Take a June 2022 decision by Judge George B. Daniels of the Southern District of New York. Judge Daniels was weighing an application from Luis Andre Delacruz, who had served a 10-year prison sentence. Delacruz had essentially met the standard expectations of his release: he held a full-time job, stayed sober, and lived with his family. Probation had no objection, and the order records no objection from the government.
Pointing the AI agent at this document extracts the judge’s logic immediately: Judge Daniels granted the application in a single page, wiping away nearly three years of supervision. He required no extraordinary change in circumstances and cited no legal precedent. Instead, the order simply rests on Probation’s report of basic compliance: no arrests, clean drug tests, and supervision served “without incident.”
This logic is not isolated. Scanning cases across the country, the AI found Judge Dale A. Drozd applying the same standard in the Eastern District of California. In United States v. Hill, Judge Drozd granted early termination to a defendant whose case rested on a clean record, a stable life, and the non-objection of Probation and prosecutors. Under this view, the rule is straightforward: basic compliance satisfies the “interest of justice.”
But let the AI read further, and a contradictory legal principle emerges. In 2023, Judge Gregory H. Woods—sitting in the same district as Judge Daniels—reviewed an application from Christopher Gay. While Gay had a prior supervision violation, he had since maintained steady work and performed community service. Judge Woods explicitly dismissed those achievements. Quoting earlier case law, he ruled that such conduct is simply “what is expected of a person under…supervised release,” not a reason to end it early.
That strict view traces back to a 1997 appellate decision, United States v. Lussier, which noted that “occasionally, changed circumstances—for instance, exceptionally good behavior”—could justify altering a sentence. Over time, many judges have hardened that observation into a rigid rule: without exceptionally good behavior, there is no early termination.
The appeals courts, however, have rejected that rigid rule. In 2016, the Second Circuit — the same court that issued Lussier—clarified in United States v. Parisi that its earlier decision “does not require new or changed circumstances” but “simply recognizes that changed circumstances may in some instances justify a modification.” Four years later, the Third Circuit in United States v. Melvin called the strict interpretation a flat-out “misreading of Lussier,” stating the holding plainly:
“We therefore hold that a district court need not find that an exceptional, extraordinary, new, or unforeseen circumstance warrants early termination of a term of supervised release before granting a motion under 18 U.S.C. § 3583(e)(1).”
An AI agent can trace this citation trail in a matter of seconds, cutting through the confusion to extract the appellate position. The appeals courts that have addressed it agree: extraordinary circumstances might be sufficient, but they are not necessary.
When judges follow the appellate courts, the outcomes shift entirely. In the 2022 New Jersey case United States v. Hawkins, the defendant had a steady job, deep church involvement, and plans to marry. Probation did not object, but the government did, deploying the same “compliance is merely expected” argument Judge Woods adopted. Judge Katharine S. Hayden, relying on Melvin, held that Hawkins did not need “exceptional, extraordinary, new, or unforeseen circumstances.” She found that he had “done more than merely comply.” Calling his day-to-day achievements “quietly compelling,” and declaring that his sentence must be “sufficient, but not greater than necessary,” she granted early termination over the government’s objection.
When AI scales this analysis across more decisions, it reveals a stark judicial split. Some judges treat “exceptionally good behavior” as a firm requirement, denying well-adjusted individuals. Other judges follow the appeals courts and reward basic, successful adjustment. If the governing legal standard changes depending on which judge happens to catch the case, it shatters the illusion of equal treatment under the law.
Let us return to the cliffhanger at the end of Part 3. Imagine a defendant who completed an intensive reentry program, was so successful at work that he was promoted to store manager, stayed sober, and sat on the lowest level of supervision. He asked the judge to end his supervision just 73 days early. The government objected. If you wrote down your ruling after reading Part 3, now is the time to check it.
The judge who caught that case was Judge Daniels—the same judge who granted Delacruz’s release in 2022. By April 2026, Hector Tirado’s record of adjustment was arguably far stronger than Delacruz’s. Tirado had not only maintained basic compliance; he had completed the district’s demanding RISE reentry court, worked 48 hours a week plus weekend security shifts, and stayed entirely clean. Aside from minor traffic offenses, he had fully complied. He was asking to wipe away less than three months of supervision.
Based on the previous ruling in Delacruz, we would expect Judge Daniels to easily grant Tirado’s request to wipe away those final 73 days. Instead, in a five-page denial, he abruptly switched legal standards. He dismissed Tirado’s accomplishments as simply fulfilling expectations. Pointing the AI at the order extracts the logic in one stark sentence based on Lussier:
“Defendant’s motion for early termination of his term of supervised release is DENIED because he has not demonstrated extraordinarily good behavior or unforeseen circumstances which justify early termination.”
Judge Daniels did note the seriousness of Tirado’s underlying offense and a prior sentence reduction. Yet his governing standard was the rigid reading of Lussier—a “misreading,” according to the Third Circuit.
None of this is a criticism of Judge Daniels or any other judge. Deciding whether supervision has served its purpose is exceedingly difficult: it requires weighing many variables against a vague standard and combining them into a single, binary decision. Identifying those variables, determining their magnitude, and deciding how they weigh relative to one another is simply hard. Arguably, it is the most difficult type of decision-making to study and understand.
In Noise: A Flaw in Human Judgment, Daniel Kahneman, Olivier Sibony, and Cass R. Sunstein devote an entire book to the complexities involved with these decisions. Not surprisingly, the book opens with examples pulled from the courts. They recount Judge Marvin Frankel’s dismay that similar cases received wildly different sentences, and a study in which identical hypothetical cases, put to federal judges, produced answers so varied that “absence of consensus was the norm.” They catalog stranger influences still: harsher decisions after a local team’s weekend loss, and more leniency on a defendant’s birthday.
If the best minds on the bench cannot escape this noise, inconsistency in early termination should surprise no one. What is new is that it can now be measured. AI can read every decision, record the standard applied and the facts weighed, and show where similar cases diverge. Making court documents accessible in bulk would open a floodgate of opportunities for measuring and strengthening the rule of law.
Measuring judicial noise is retrospective work. It relies on AI looking backward to recognize and compare patterns. But these same tools can be deployed while a case is still being decided, cross-checking the claims in every filing against the evidence already on the record.
Imagine a fiercely contested civil case in state court: nearly a dozen days of testimony, millions of dollars at stake. At the conclusion, the judge invites both sides to submit proposed findings of fact—a standard, fully permitted practice—and then issues his own opinion.
But do judges actually copy from those partisan proposals? To a layperson, adopting a lawyer’s exact words might look like the kind of plagiarism that gets college students expelled. Put that aside, and assume this judge copied at least one sentence. Out of every 100 sentences in the judge’s final findings, how many do you think were lifted, word-for-word or nearly so, from each side?
Pick two numbers, write them down, and save them. The answer comes in Part 5.
David Stasior, MD, MPP, is a principal at Broken Scales, where he applies data analysis to the justice system, and the author of Slot Machine Justice.
