Structural Weighting in Pretrial Risk Algorithms: How Proprietary Matrices Calculate Detention

Before a judge ever hears a bail argument, a defendant's probability of going home or going to jail is determined by a mathematical matrix. In jurisdictions using algorithmic pretrial risk assessments, a software tool ingests a defendant's criminal history, assigns point values to specific past behaviors, and outputs a risk score that dictates the baseline recommendation for release or detention. Because judges anchor their decisions to these automated outputs, the specific weights assigned to those inputs—how heavily a missed court date or a prior misdemeanor is penalized—function as the actual mechanism deciding pretrial freedom.[4]

The shift toward algorithmic justice began accelerating in 2013, when the Laura and John Arnold Foundation (now Arnold Ventures) released the Public Safety Assessment (PSA). Designed using a dataset of 1.5 million cases from approximately 300 jurisdictions, the PSA was intended to replace subjective, often biased judicial intuition with a standardized statistical model. Today, the PSA and competing proprietary tools like Equivant's COMPAS and the University of Cincinnati's ORAS-PAT are utilized in courts across the United States, assessing millions of defendants annually.[2]

These tools operate on a shared premise: historical data can predict future behavior. When a defendant is booked, an intake officer or automated system pulls their criminal record and feeds it into the algorithm. The software evaluates static factors—such as age at first arrest, prior felony convictions, and past failures to appear in court—and calculates a numerical risk score.[4]

However, the exact calculation of that score depends entirely on which vendor the jurisdiction has licensed. The Arnold PSA, for example, relies on exactly nine variables and explicitly excludes demographic data like race, employment status, and residential stability. It generates two separate scores on a scale of 1 to 6: one predicting the likelihood of a new criminal arrest (NCA) and another predicting failure to appear (FTA).[2]

In contrast, the Ohio Risk Assessment System Pretrial Assessment Tool (ORAS-PAT), developed by the University of Cincinnati Corrections Institute, utilizes a seven-item matrix that includes socioeconomic factors. It evaluates employment at the time of arrest, residential stability, and illegal drug use over the past six months, compressing these inputs into a single risk score ranging from 0 to 9.

Equivant's COMPAS system employs an even more expansive and proprietary methodology. Categorized as a fourth-generation risk assessment instrument, COMPAS evaluates 137 total questions across 15 sub-scales, including family criminality, peer associations, and substance abuse. It places defendants into risk deciles from 1 to 10, with scores of 8 to 10 labeled as "High" risk.[1][3]

The structural divergence between these matrices means that the exact same defendant will receive vastly different detention recommendations depending on the county in which they are arrested. A central point of divergence is how the algorithms penalize a prior missed court date.[4]

Under the Arnold PSA, a prior failure to appear within the past two years adds 1 or 2 raw points to the defendant's FTA score, which is then converted to the 1 to 6 scale. A single missed appearance increases the risk profile, but the segmented scoring ensures it does not automatically inflate the defendant's predicted risk of committing a new violent crime.[2]

Conversely, tools that compress all risk factors into a single composite score amplify the penalty of a missed court date. Because the ORAS-PAT merges flight risk and recidivism risk into a single 0 to 9 scale, a defendant who missed a court date due to lack of transportation two years ago receives a higher overall risk classification, which a judge may interpret as a generalized threat to public safety.

The opacity of proprietary algorithms further complicates the judicial process. While the Arnold PSA's scoring weights are publicly available, the exact mathematical weights used by Equivant's COMPAS remain a trade secret. As Equivant (formerly Northpointe) noted in its practitioner's guide, "scores in the medium and high range garner more interest from supervision agencies than low scores, as a low score would suggest there is little risk of general recidivism."[1]

This proprietary opacity became the subject of national scrutiny in May 2016, when a ProPublica investigation analyzed the COMPAS scores of 11,757 defendants in Broward County, Florida. The analysis revealed that the algorithm's overall predictive accuracy for recidivism was roughly 62.5 percent for white defendants and 62.3 percent for Black defendants.[1]

However, the distribution of errors was heavily skewed. ProPublica found that Black defendants were 45 percent more likely to be assigned a higher risk score than white defendants who went on to commit similar future crimes. The algorithm falsely flagged Black defendants as future criminals at almost twice the rate of white defendants (44.9 percent compared to 23.5 percent).[1]

Advocates for algorithmic risk assessments argue that the tools, while imperfect, remain significantly more objective than human judges. The National Association of Pretrial Services Agencies maintains that evidence-based practices, including validated actuarial assessments, are essential for reducing unnecessary pretrial detention and eliminating the wealth-based disparities of the cash bail system.

"The PSA's nine factors must be used to calculate three scores: Failure to Appear (FTA), New Criminal Arrest (NCA), and New Violent Criminal Arrest (NVCA)," states the Advancing Pretrial Policy and Research guide. "FTA and NCA must be reported as a scaled score, and NVCA must be reported as the presence or absence of a flag. Report these three scores separately. Do not combine them into one score."[2]

This separation of scores is a direct response to the compounding errors of earlier algorithms. By isolating the risk of flight from the risk of violence, the PSA attempts to prevent judges from detaining non-violent defendants simply because they have a history of missing court dates.[2][4]

Despite these refinements, the translation of a risk score into a legal outcome remains highly localized. Once an algorithm outputs a score, local jurisdictions apply a secondary "decision-making framework" or "release matrix." This policy document dictates whether a score of 4 results in release on recognizance, mandatory GPS ankle monitoring, or a recommendation for pretrial detention.[4]

Consequently, the algorithm itself does not set bail; it establishes the empirical baseline that the court uses to justify its decision. If a county's release matrix dictates that any defendant scoring above a 5 on the COMPAS scale must face a detention hearing, the algorithm's proprietary weighting of a prior misdemeanor conviction effectively becomes the deciding factor in whether that individual retains their freedom.[4]

As jurisdictions continue to adopt these tools to comply with bail reform mandates, the mathematical mechanics of risk assessment will increasingly dictate the boundaries of the American pretrial system. The specific variables an algorithm includes—and the exact mathematical weight it assigns to them—will determine the liberty of millions of defendants long before a judge strikes the gavel.[4]