A Stable Attribution Is Not a Stable Error Rate: What Feature-Attribution Monitoring Can See Under Drift, Why 'More Sensitive Than the Inputs' and 'Silent While Fairness Fails' Are Both Correct, and the Label-Free Statistic the Fairness Alarm Needed

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Deployed classifiers are often monitored without labels, and feature attributions are one of the signals watched: a production monitoring service alerts when the ranking of SHAP attributions moves, and one line of work argues that the explanation space is more sensitive than the input space to shifts that matter. A 2026 study reports the other face: on COMPAS under injected feature noise, TreeSHAP explanations stayed largely stable while the protected group's false positive rate rose from 37.84 to 70.39 percent. In this analysis the author reviews and synthesises published results to show that the two findings concern different objects. Per-instance attributions, the distribution of attribution vectors, a top-k ranking of mean absolute attributions and an attribution of the loss are four instruments; only the last uses labels, and only the second sees everything the prediction distribution sees. The paper orders the label-free observables into a ladder, derives which group fairness quantities each rung can read, and shows that a group false positive rate is readable from none of them without an assumption about how outcomes given features changed. The anchor study's own numbers add two results: its disparity widened as a difference and narrowed as a ratio, and, with labels fixed, its false-positive rises imply rises in each group's positive-prediction rate, which needs no labels to compute, unless a group has almost no true negatives. The paper gives a decision procedure for choosing and reading a label-free fairness monitor and names the studies that would settle what remains open.