Algorithm Validation as a Policy Audit: Evidence from Race-blind Charging

TEXT START: California recently required all prosecutors in the state to conduct a "race-blind charging" decision by reviewing case documents in which selected race-related proxies have been redacted.

THE DISSECTION

The paper performs two audits: whether bc2 obeys California’s mandate, and whether the mandate achieves race-blind decision-making. Its central finding is the gap between them: 96.7% technical compliance still leaves major racial proxies—especially location—untouched. Additional redaction removes 43.1% of the remaining predictive signal.

The real move is to convert a political and institutional problem into a measurable software pipeline: proxy lists, document transformations, accuracy rates, and iterative patches. That makes algorithmic governance appear controllable while handing the definition of relevant information to model designers and validators.

THE CORE FALLACY

The paper’s evidence partly destroys the policy’s premise. Removing textual proxies is not the same as removing racialized decision-making. Police deployment, geography, reporting practices, socioeconomic conditions, and prosecutorial inference remain embedded in the case material and institutional process.

A 43.1% reduction in predictive signal is not a 43.1% reduction in discrimination. The signal may contain illegitimate racial proxies, legitimate operational context, or both. Likewise, document-level compliance does not establish improved charging outcomes. The policy can be faithfully implemented and still fail substantively—which is precisely the paper’s most damaging result.

Under the Discontinuity Thesis, this is a lag defense: cognitive automation preserves procedural legitimacy while shifting governance into the machinery of redaction, validation, and metric selection.

HIDDEN ASSUMPTIONS

• Race can be represented by a finite, knowable list of textual proxies.
• Redacting proxies prevents prosecutors from reconstructing race from remaining context.
• Predictive association is equivalent to discriminatory contamination.
• Document-level accuracy predicts case-level fairness and charging outcomes.
• Nearly 5,000 reports from jurisdictions across the United States generalize to California’s deployment.
• The model, proxy definitions, and legal mandate will remain stable rather than being gamed or revised.
• Removing additional information improves neutrality without damaging legitimate decision quality.
• Validation can correct the policy without confronting the underlying institutional production of the data.

SOCIAL FUNCTION

Primary classification: partial truth and transition management. Secondary classification: prestige signaling and institutional self-exoneration.

This is not pure copium. It identifies a real failure in the mandate and demonstrates that compliance is not success. But it also makes a legitimacy crisis administratively digestible: if the problem can be assigned to missing proxies or an earlier model version, institutions can claim progress without surrendering the broader system of discretion.

THE VERDICT

Useful as a narrow audit. Inadequate as evidence of race-blind justice.

The paper improves a redaction machine, not the unequal system producing police reports or the human institutions interpreting them. Its durable contribution is exposing the gap between procedural compliance and substantive neutrality. Under DT logic, bc2 is a fairness wrapper around expanding cognitive automation: it may delay institutional failure, but it cannot reverse the transfer of judgment into algorithmic infrastructure. It finds a leak in the ship and mistakes the repair log for salvation.