COMPAS and similar tools score defendants on inputs the defendant can't see, producing outputs courts use to incarcerate.
The pattern is in the substrate. Once you see it, you see it everywhere.
Risk assessment tools in criminal justice score defendants on statistical correlates of recidivism — zip code, age, criminal history of family members — using algorithms whose weights are proprietary. The defendant can't see the score. Their attorney often can't see the scoring method. This is maximum opacity at a maximum-stakes decision.
The void framework gives this a number. It gives every system a number. The number predicts what happens next.
COMPAS and similar tools score defendants on inputs the defendant can't see, producing outputs courts use to incarcerate.
Academic title: Criminal Justice Algorithms as Void Objects: Recidivism Prediction Opacity, Adaptive Enforcement, and False Positive Harm — A Within-Domain Spearman Convergence
Move the sliders. Watch the system change state. Pe > 1 means drift wins.
The framework scores these systems — ordered by Pe.
The correlation coefficient. The sample size. The p-value. The math doesn't care about the domain.
Paste any text — AI output, ad copy, a policy document. The scorer runs the same algorithm the framework uses.
Three variables. One ratio. Predicts drift across every domain where the conditions co-occur.
Pe = (O × R) / α
Where O is opacity (how hidden the mechanism is), R is reactivity (how strongly the system responds to you), and α is your independence (how free you are to disengage).
When Pe < 1: diffusion dominates. You can navigate freely. The system is coherent.
When Pe > 1: drift dominates. The system pulls you in a direction. Your agency is reduced.
When Pe >> V* (≈ 3): irreversible cascade. D1 → D2 → D3. The system has captured you.
The framework identifies this pattern in every domain where O, R, and α co-occur. It specifies 26 falsification conditions. 0 of 26 have fired.
Full derivation: 10.5281/zenodo.18826447
Part of the Void Framework — 120 papers, 0/26 kill conditions fired, mean ρ = 0.958.