Predict-then-optimize pipelines make decisions by first predicting an unknown outcome and then optimizing against that prediction. A new arXiv paper notes that any fairness constraint in such a pipeline must be applied to the predicted outcomes, because the true outcomes are not available when the decision is made.

The paper asks a pointed question: if you enforce demographic parity on those predicted outcomes, does that actually translate into demographic parity for the true outcomes? The abstract frames this as the central issue but, in the excerpt provided, does not reveal the answer.

This is a useful reminder that fairness guarantees are only as meaningful as the quantities they are computed on. Without knowing the paper's findings, readers should treat the abstract as an open question rather than a settled result.