Person-level sentencing data, across states
There is no person-level, multi-state record of how the United States sentences people in its state courts. Just Scales is building one — so advocates, lawyers, and legislators can compare a sentence against comparable cases and against the law in force when it was imposed.
As of 2026. FBI NIBRS arrest linkage and Census ACS demographic overlay are in progress.
Mission
“Verifiable data, presented with full context, so that anyone can see what is happening and decide for themselves what it means.”
Just Scales follows the data at the individual level, wherever it leads. It exists to support equal application of the law — appropriate consequences applied fairly — not blanket leniency. Every pattern it identifies is a hypothesis for review, not a verdict about a person.
The problem
Who sentences a defendant, whether they can post bail, which county charges them, and what year they were convicted all change the outcome — often more than the conduct itself. The evidence is documented in the research, but it cannot be acted on case by case, because the underlying records are fragmented across 50 incompatible state systems.
Defendants randomly assigned to different judges in the same courthouse receive different sentences for the same offense. After United States v. Booker made the federal guidelines advisory, inter-judge disparity approximately doubled. In the District of Massachusetts, the gap between the most lenient and most severe judge grew from 15 months to nearly 40 months after Booker.
Yang, NYU Law Review, 2014; Scott, Stanford Law Review, 2010
Defendants detained before trial because they cannot post bail are convicted at higher rates, primarily through guilty pleas. Pretrial detention increased the probability of pleading guilty by 10.8 percentage points. Detained misdemeanor defendants in Harris County, Texas were 25% more likely to plead guilty and 43% more likely to be sentenced to jail, after controlling for charge, demographics, and criminal history.
Dobbie, Goldin & Yang, American Economic Review, 2018; Heaton, Mayson & Stevenson, Stanford Law Review, 2017
In randomly assigned Philadelphia murder cases, defendants represented by public defenders were 19% less likely to be convicted of murder and 62% less likely to receive a life sentence than those assigned court-appointed private attorneys. About 73% of county-based public defender offices lack sufficient staffing; the difference was institutional — appointed attorneys had no investigators, office support, or co-counsel.
Anderson & Heaton, Yale Law Journal, 2012; RAND National Public Defense Workload Study, 2023
Initial charging decisions — particularly whether to file charges carrying mandatory minimums — explain more variation in sentence outcomes than judicial sentencing. Constraining judicial discretion does not reduce disparity; it transfers control to prosecutors, whose decisions are less visible and less subject to review.
Starr & Rehavi, Yale Law Journal, 2013; USSC Report to Congress, 1991
Imprisonment rates range from 153 per 100,000 (Maine) to 816 per 100,000 (Louisiana). An analysis of more than 500,000 federal sentencing records (2006–2020) found Black defendants received sentences about 19 months longer at the system-wide level, with 14 districts showing differential treatment after controlling for offense, criminal history, and demographics.
Bureau of Justice Statistics, Prisoners 2014; Humanities & Social Sciences Communications, 2023
People remain incarcerated for conduct the legislature has since decided does not warrant their sentence. After California's Proposition 36 (2012), more than 2,200 people were resentenced; one had received 50 years for shoplifting $153 of videotapes. Washington removed second-degree robbery as a three-strikes offense in 2019 but did not make the change retroactive; 62 people remain serving life for that offense. Of 11 states that reformed three-strikes laws, only California applied the reform retroactively.
California Policy Lab; Courthouse News
The scale
About 1.9 million people are incarcerated (Bureau of Justice Statistics, 2022). Hundreds of thousands may be serving sentences that do not reflect current law or comparable cases. These categories overlap, but no one can say by how much, because the data does not exist in linked form.
| Category | Estimate | Source |
|---|---|---|
| Currently incarcerated | 1,900,000 | BJS, 2022 |
| Sentenced under later-amended law | 200,000–400,000 | Internal estimate |
| Geographic outlier sentences | 100,000–300,000 | Internal estimate |
| Clemency / resentencing eligible | 50,000–150,000 | Internal estimate |
Internal estimates, not independently verified.
Why it persists
More than 252,000 defense lawyers and dozens of innocence and second-look organizations work these cases. The number who can acquire, clean, and analyze multi-state incarceration data is near zero. Case-by-case analysis costs $100,000–$500,000.
States differ not only in charges but in sentencing frameworks, offense categories, and data formats — from documented CSVs to fixed-width mainframe dumps to scanned PDFs. Comparing a drug sentence in Texas to the same charge in Florida requires mapping different agencies, formats, and laws into one model.
Nine organizations in the field hold more than $750M in combined assets and $640M in combined revenue (IRS Form 990 / ProPublica). None has built person-level, multi-state sentencing data. Government has not either: the Bureau of Justice Statistics publishes aggregate data with multi-year delays; the U.S. Sentencing Commission covers federal courts only.
The solution
Solution 1
A single queryable system of individual sentencing records from state Departments of Corrections, standardized across 15+ dimensions and linked to FBI arrest data and Census demographics. For the first time, a sentence in one state can be compared to the same charge in another.
State by state: identify the agencies that hold the records, file public-records requests, and handle incompatible formats — PDFs, mainframe exports, spreadsheets, headerless CSVs. Then clean, standardize, and load into a unified model with full data lineage, and overlay FBI NIBRS arrest data and Census ACS demographics.
Ranked lists of anomalous sentences; cohorts matching resentencing or clemency criteria; judge-versus-judge and county-versus-county comparison; pre/post-reform impact measurement. Case-specific outlier analysis that previously cost $100,000–$500,000 becomes routine.
As of 2026. FBI NIBRS linkage and Census ACS overlay are in progress.
The warehouse can show that two sentences differ. It cannot show whether a sentence was consistent with the law in force when it was imposed. That requires a second layer.
Solution 2
A lawful difference and a biased one can produce the same statistical pattern. Two people convicted under the same offense code may receive different sentences because the law required it — a crime against a child, or a stricter law in force at the time — or because of bias. In the records alone, these look identical. SLSL reconstructs what the law required for a given offense, in a given state, on a given date.
The person-level record from DOC data; includes offense code and date.
The statutory citation (for example, ORS 163.405) identifying the specific crime.
The severity grouping (for example, Crime Seriousness 1–11); offense codes move between categories over time.
The legislative documents defining what the category meant at that moment, including priors, enhancements, and mandatory minimums.
The minimum and maximum, departure allowances, and supervision terms for that offense at that point in history.
Sentencing law is not centralized, is not structured for analysis, and changes independently across many dimensions. States replace entire frameworks — Oregon has had four since 1960 — and when a framework changes, the categories and rules change meaning. No prior source links sentencing legislation to individual records over time.
Status (in development as of 2026)
Legislative histories are in progress across 50 states, resolving 2,000–5,000 sentencing-relevant events. Where the lineage is available, each record can be paired with the sentence range the law set on its date.
Principles
Justice for the individual regardless of group. If a trend fits a particular pattern, it is reported; if it does not, that is reported too.
Work is directed by the number of people affected, the life-years at stake, and the likelihood of relief.
The goal is consequences applied fairly, not arguing people out of prison.
Data is presented with legislative, demographic, arrest, and geographic context, with full lineage, documentation, and code availability so results can be reproduced.
Validation
“I've been doing this a long time and I have no idea how you've gotten that data.”
“It blows my mind it doesn't exist… really brilliant.”
“I would love something like this when we are writing clemency petitions.”
“A very user-friendly way of setting up the data… very appealing and exciting.”
“It's unbelievable you guys would do this on your own.”
Also reviewed by Dwight Holton (former U.S. Attorney), Craig Prins (Oregon CJ Commission), Nicholas Sensley, Rep. Karen Dolan, Sean Litton (Technology Coalition), Tricia Bushnell (Quattrone Center, University of Pennsylvania), and the Villanova Law Library.
Team
Just Scales operates as a distinct, not-for-profit initiative of Agents & Allies, which holds SOC-2 Type I and II and also operates Schemantic.io. The team has conducted 14+ domain-expert interviews and acquired 21 states of data.
CEO, Agents & Allies / Storytellers.ai
Former Amazon: Head of Marketing for Alexa Shopping and Marketplace. Led 14+ criminal-justice domain-expert interviews and the acquisition of 21 states of data.
CTO, Agents & Allies; CEO, Schemantic.io
Former Amazon: nine-year veteran who led a core science team and managed engineering, science, and economics groups at global scale. University of Washington faculty.
Head of Engineering, Schemantic.io / Agents & Allies
Former Amazon: Marketplace and AWS. National Merit Scholar. Builds the data-acquisition and transformation pipeline.
Data Scientist, Agents & Allies
Yale. 15+ years at the intersection of code and language; former DOE ARPA-E attorney-advisor. Published Supreme Court datasets on Kaggle. Leads machine learning.
Program Manager, Just Scales
Villanova: Political Science, Peace & Justice (2025). Manages public-records requests and data acquisition across 21 states.
Just Scales is structured separately from the for-profit Agents & Allies for legal protection, focus, and mission durability. Agents & Allies was recognized among Inc. Magazine's fastest-growing companies (Pacific #120, 2024).
Support
The field holds more than $750M in combined assets and $640M in combined revenue, yet no person-level, multi-state sentencing data has been built. Just Scales is asking for about 0.26% of what the field holds to build what it has named as its most pressing gap.
Use of funds
| Category | Share | Detail |
|---|---|---|
| Data acquisition | 30% | Expand from 21 to 35 states. |
| Engineering | 30% | Census ACS overlay, FBI NIBRS linkage, ML cohort analysis, legislative-lineage attachment. |
| Operations | 20% | Public-records processing, data hygiene, documentation. |
| Partnerships / outreach | 15% | 50+ partner analyses; lawyer outreach. |
| Grant infrastructure | 5% | 50+ annual applications toward a $1M target. |
Roadmap
The dataset could be commercialized the way legal-research products restrict access to those who can pay. Just Scales has chosen not to. For those using the data to do good, access is free.
Contact
Funders
To discuss the $2.5M and the path to sustainability, contact Wing Yew Lum.
Email Wing Yew LumAdvocates, lawyers & researchers
Public defenders, legal-aid and innocence organizations, qualifying law-school clinics, and researchers can request access to tools and data views beyond the public layer. Partners receive secure, credentialed access appropriate to their work.
Request accessOr email contact@justscales.org.
