COMPAS vs SAVRY (Structured Assessment of Violence Risk in Youth)

COMPAS excels at generating a consistent, statistically derived risk score based on a broad set of 137 static and dynamic factors. Its strength lies in its actuarial purity—removing human subjectivity to produce a recidivism probability. For example, a 2016 ProPublica analysis found COMPAS correctly predicted recidivism roughly 61% of the time, but its design is normed on adult populations where criminal history is the dominant predictor, ignoring the developmental malleability central to youth justice.

SAVRY takes a fundamentally different approach by anchoring risk assessment in Structured Professional Judgment (SPJ). Instead of a black-box algorithm, it guides clinicians to evaluate 24 risk factors and 6 protective factors specifically validated for adolescents. This results in a trade-off: SAVRY introduces clinical oversight that can account for developmental maturity and environmental interventions, but it sacrifices the mechanical consistency of an actuarial tool like COMPAS. Meta-analyses show SAVRY's summary risk ratings achieve median AUCs of 0.71 to 0.78 for violent reoffending, comparable to adult tools but with a youth-specific context.

The key trade-off: If your priority is pure statistical prediction and automation of a risk score without human bias, COMPAS offers a consistent, albeit adult-normed, output. If you prioritize developmental sensitivity, the integration of protective factors, and a structured clinical override that aligns with juvenile justice philosophy, SAVRY is the legally defensible standard. Applying COMPAS to youth courts risks violating the constitutional requirement for individualized sentencing by ignoring the transient immaturity of adolescence.

COMPAS excels at generating a rapid, statistically derived risk score from a large dataset of primarily adult offenders. Its strength lies in its consistency and scalability; a 2016 ProPublica analysis highlighted its ability to produce uniform scores, though it also revealed significant racial disparities in false positive rates for Black defendants. For a CTO, this represents a system with deterministic, low-latency outputs, but one that carries substantial 'algorithmic debt' in the form of baked-in societal biases that are difficult to audit or correct without a full model rebuild.

SAVRY takes a fundamentally different architectural approach by mandating Structured Professional Judgment (SPJ). Instead of a black-box score, it guides a clinician to weigh historical factors against dynamic, youth-specific items like 'peer rejection' and 'community disorganization,' while crucially integrating protective factors such as 'strong attachment to school.' A meta-analysis by Singh et al. (2011) found SAVRY summary risk ratings achieved a median area under the curve (AUC) of 0.71 to 0.78 for predicting violent reoffending, demonstrating strong predictive validity without sacrificing the contextual nuance of adolescent development.

The key trade-off centers on the definition of 'ground truth.' COMPAS optimizes for statistical efficiency and inter-rater reliability by removing human discretion, which is valuable if your priority is a fast, automated triage system for a high-volume adult system. SAVRY optimizes for developmental validity and intervention planning, accepting the cost of human administration to gain a dynamic, case-formulated risk picture. If your juvenile justice system prioritizes rehabilitation and needs to identify malleable treatment targets, SAVRY's SPJ model is the ethically and operationally superior choice. Choose COMPAS only if your use case is strictly limited to static, historical risk classification in an adult context where developmental factors are irrelevant.