The next discovery came from a feature that appeared harmless: postal zone.
The original developers had included it because travel distance, neighbourhood resources, and school access varied geographically.
Postal zone was not race, income, disability, or family status.
Yet when the team mapped predictions, the variable reproduced several historical patterns of segregation with remarkable fidelity.
Priya described the feature as a proxy: a variable that indirectly carries information about another attribute.
Removing an explicitly sensitive attribute therefore does not guarantee that a model is blind to it.
Other variables can reconstruct much of the same information through correlation.
Elias tested this by training a separate model to predict household income using only features CivicLearn already contained.
The result was disturbingly accurate.
Internet stability, commute time, school transfer frequency, housing turnover, and extracurricular participation together formed a socioeconomic signature.
No single feature revealed income directly.
The combination did.
This created a difficult design question.
Should the team remove every variable correlated with social disadvantage?
Doing so might reduce discriminatory effects, but it might also erase information necessary to identify students who genuinely needed support.
A system designed to allocate assistance cannot become useful by pretending inequality does not exist.
Priya distinguished between using sensitive information to discriminate and using it to detect or correct unequal outcomes.
The same variable can be harmful in one decision context and necessary in another.
Context determines function.
The team experimented with constrained models that limited reliance on certain proxies while preserving predictive value.
They also compared outcomes with and without sensitive attributes included explicitly for auditing.
Elias had once assumed that ethical design meant deleting controversial variables.
He now understood that omission can reduce visibility without reducing influence.
Sometimes the only way to discover discrimination is to measure the categories through which discrimination occurs.
The proxy problem destabilized the intuitive notion that fairness can be achieved through simple data blindness, because structural inequality leaves redundant statistical signatures across many variables.
Selective omission could therefore become performative innocence: the institution could claim not to 'see' a category while continuing to act on information that reconstructed it indirectly.
The proxy problem therefore turned privacy, fairness, and transparency into competing objectives rather than isolated requirements.