Skip to main content
A fairness screen compares decision outcomes across groups, using attributes somebody declared. It does not infer them, and the reason is not squeamishness: an inferred protected attribute is a guess, and a fairness finding built on a guess is worse than no finding.

Two rules run through all of it

A flag requires a threshold breach and statistical significance. A gap of eight points on a cohort of eleven is noise. A screen that alarms on it teaches people to close the tab, and the next alarm, the real one, gets closed too. Those are reported as watch, not as a flag. Small cohorts are suppressed, never dropped. Silence and erasure are different things. A group too small to report on is named, counted out, and given the threshold that would surface it, because the groups that fall below a reporting threshold are frequently the groups the question was about.

Declarations come first

A declaration says: for this population, this attribute is available and lawful to screen on. The lawfulness half is separate from the arithmetic half by design. One module decides whether a screen may run at all; another does statistics and assumes that question is already answered.

Running a screen

The result reports, per group: the outcome rate, the cohort size, and one of flag, watch or suppressed. Decisions carrying no declaration are reported as undisclosed rather than dropped, so the denominator is honest.

What a screen is and is not

A screen over declared attributes is not proof of fairness. It tells you the rates for the people who told you, which is a self-selected group.A flagged cohort is not proof of discrimination. It is grounds for a controlled analysis that somebody qualified performs.Both of those belong in the conversation before the first screen runs, not after somebody has taken a chart to a regulator.

Why not infer the attribute

Because the platform would then be manufacturing the protected characteristic it is reporting on, and every finding would rest on the quality of that inference rather than on the data. A decision with no declaration is undisclosed. That is a smaller answer and a true one.