Use case · Criminal justice
COMPAS: the canonical case of racial bias encoded in data.
COMPAS is a commercial tool that scored criminal defendants' likelihood of reoffending, used in US courtrooms in bail and sentencing decisions. It is the most studied example of algorithmic racial bias in the world, which makes it the right public dataset to demonstrate on.
The problem, as ProPublica published it
In 2016, ProPublica investigated COMPAS scores for thousands of defendants in Broward County, Florida. Its published finding: the tool falsely flagged Black defendants as future criminals at roughly twice the rate of White defendants, while White defendants who did go on to reoffend were more often mislabelled low-risk.
The race column is not the mechanism. The bias rides in through proxies: prior counts, charge categories, age interactions, and other features that correlate with race in the underlying data. Any model trained on that data learns the prejudice along with the signal.
Source: ProPublica, "Machine Bias" (2016), cited here as public fact.
What Rosa does about it
Rosa, applied to ProPublica's two-year COMPAS cohort with race as the protected attribute, narrows the false-positive-rate gap by race in a downstream recidivism model - ProPublica's own disparity metric - by about 81% (mean, range 54-100% across 10 independent Rosa runs), with a demographic-parity-gap reduction of about 90%, while preserving each column's statistics. COMPAS is the deliberately hard, weak-signal case, and we show the honest cost: the debiased model's overall accuracy falls by about 14 points, toward the majority-class baseline, because the raw model was leaning heavily on race - so removing race's influence takes much of its ranking power with it. On data with genuine non-protected signal there is no such trade-off (see the credit experiment on the Validation page). The FAN bias dial falls from 0.126 to 0.009 on the same runs. These are two different bias metrics, and Rosa leads with the tougher one: the 81% is the stricter, thresholded false-positive-rate gap. The white paper below reports a different, distribution-level measure - the standardised gap in the model's mean risk scores - which Rosa drives to near-zero, consistent with the dataset bias dial falling from 0.126 to 0.009 above.
That is a reproducible result on a public dataset: a buyer can verify it with the product being sold. It is the difference between citing a famous harm and showing the fix.
The published Demonstrating Rosa white paper shows the same pattern at the model level: a recidivism model trained on the raw ProPublica data scored Black defendants higher whether or not they went on to reoffend (standardised bias 0.85 for non-recidivists, 0.92 for recidivists). Trained on Rosa-debiased data, the gap effectively vanished (0.00 and 0.17). Read the white paper (PDF).
The most studied case of bias in data
ProPublica: COMPAS flagged Black defendants as future criminals at roughly twice the rate of White defendants.
gap by race, removed
downstream recidivism model
Who this is for
Public sector, justice, and any organisation running high-stakes scoring of people. Under the EU AI Act, law enforcement uses are named high-risk, which puts the training data squarely under Article 10's examination and mitigation duties. This page doubles as Rosa's racial-bias case: the mechanics are identical wherever race proxies live in your data.
Your data has its own COMPAS story. Measure it.
Diagnose reports your measured bias and the per-column proxies behind it.