Passage A:

Pretrial risk scores estimate the likelihood that a person released before trial will miss court or be rearrested. Defenders claim that a numerical score disciplines intuition. But a score can conceal policy judgments more effectively than a judge's stated reasons. Designers choose the outcome to predict, the time horizon, and the cost assigned to false positives. A model trained on rearrest data also learns policing patterns, not simply conduct, because arrest depends on surveillance and enforcement.

Due process therefore requires contestability. A defendant should receive the variables used, the score's validation results for relevant populations, and an opportunity to correct data and challenge the inference connecting score to detention. Proprietary source code need not always be public; many objections concern inputs and validation rather than code. But the government cannot justify confinement with a number whose meaning and limitations the affected person is forbidden to examine.

Passage B:

Contestability is essential, yet Passage A risks turning a screening hearing into a trial about statistical methodology. Pretrial decisions must be made quickly, often with incomplete information. Requiring litigation over every model choice may push judges back toward unaided impressions, which are neither validated nor consistently explained.

The process should separate individual error from system design. Every defendant should be able to correct personal data and show why a general factor is misleading in the particular case. Broader challenges to validation, outcome definitions, or disparate error rates should be heard periodically in institutional review proceedings with technical expertise and public participation. A hearing judge would then use only models that passed that review and would state how the score affected the decision. This division preserves meaningful scrutiny without making access to a coherent tool depend on which defendant can finance an expert during a hurried hearing.

The division of forums also requires a feedback mechanism. Hearing judges will observe recurring mismatches between a validated population model and individual cases; institutional reviewers need access to those observations when deciding whether validation remains current. Defendants and researchers should be able to trigger review with evidence of drift or unequal error, while frivolous objections may be consolidated. Without feedback, periodic approval becomes a one-time seal. Without consolidation, review may reproduce the delay Passage B fears. Durable contestability therefore depends on movement of evidence between the individual and systemic levels.

Disclosure must also be intelligible. A technical report posted online may satisfy formal transparency while giving a defendant no usable account of the variables or error rates relevant to detention. Passage A would require explanation at the individual hearing. Passage B could permit standardized explanations produced through institutional review, provided the judge connects them to the particular decision. Both approaches reject disclosure that is available in name but unusable in practice. They differ over where the underlying methodological contest should occur and how much of it a hearing can absorb without sacrificing timely adjudication.