Same Model, Different People
A fairness audit treats a trained model as a fixed decision function, but the arithmetic that evaluates it is part of that function. This holds the model, the records, the threshold and the audit code fixed and changes only the numerical execution contract: how products and sums are computed, then reduced-precision serving with BF16, FP8 and 8-bit weights.
Three tabular model families and two open-weight language models, four public tasks, CPUs and GPUs from NVIDIA, AMD and Apple. Served in FP8, the 24B model selected 54% of people instead of 45% and passed an employment audit by sex that its strict FP32 reference failed. The paper proposes an audit stability protocol: declare the contracts before scoring and report each verdict's range over them.