When the audit reached fairness metrics, Elias expected mathematics to provide clarity.
Instead, mathematics revealed incompatible definitions.
One measure asked whether students from different groups were flagged at similar rates.
Another asked whether the proportion of flagged students who truly needed support was similar across groups.
A third compared the rate at which genuinely vulnerable students were successfully identified.
A fourth focused on false positives, asking whether some groups were more often classified as needing intensive support when they did not.
These criteria could not always be satisfied simultaneously, especially when underlying outcome rates differed.
Elias found this frustrating because he had imagined fairness as a technical constraint that could be optimized alongside accuracy.
Priya argued that fairness is partly mathematical and partly normative.
The choice of metric depends on what kind of error the institution considers most harmful.
If missing a vulnerable student is the greatest concern, sensitivity may deserve priority.
If unnecessary intervention carries stigma or consumes scarce resources, false positives matter more.
Different stakeholders may reasonably disagree about which harm deserves emphasis.
The team invited teachers, students, parents, and program administrators to discuss these trade-offs.
The conversation exposed assumptions that had never appeared in the codebase.
Students worried that being labelled 'high risk' might change how adults perceived them even if the intervention was supportive.
Teachers worried that too many alerts would produce fatigue and cause serious cases to be ignored.
Administrators worried about capacity because identifying more need did not automatically create more counsellors or tutors.
Parents asked who could challenge a prediction and how long the label would remain attached to a student.
Fairness therefore expanded from a metric to a governance problem.
The team decided that no single number would be presented as proof of fairness.
Instead, they would report several metrics, disaggregated by relevant groups, and explain the value judgments behind each.
Mathematics had not removed the ethical debate.
Fairness metrics therefore operated less like neutral thermometers than competing moral lenses, each foregrounding one distribution of error while backgrounding another.
The impossibility of satisfying every criterion simultaneously did not prove fairness unattainable; it proved that fairness required an explicit account of which harms the institution was prepared to prioritize.
It had made the hidden structure of the debate harder to ignore.