Matched-calibration of changepoint detectors for biopharmaceutical quality monitoring

Monitoring a critical quality attribute (CQA) for process drift requires a changepoint detector, and non-Gaussian CQA data are widely taken to require a robust one. That recommendation is usually implemented by substituting a robust scale into an unchanged squared-error criterion. We show this implementation is the problem, not robustness itself. We compare detectors at a matched false-positive rate and a matched search-space ablation, across sample sizes 25 to 1000 and three noise distributions. At default settings each penalty’s achieved rate drifts with sample size, so uncontrolled comparison is uninterpretable. Once matched, a penalty scaled by the median absolute deviation collapses under heavy tails, its denominator failing to follow the outliers that inflate its numerator. Against an exactly known variance, that lost coupling explains a fifth to a third of the deficit under heavy tails and almost none under skew; the unstable robust scale explains the rest. A minimum segment length of five, defensible on applied grounds, removes about a third of that deficit, leaving 13 points at \(n=200\) . A bounded-loss detector shows no such collapse and, from 50 to 200 observations, leads by 15 to 22 points under \(t_3\) and 7 to 18 under log-normal, though not at 25. Where robustness enters matters more than whether it is applied. A matched ablation attributes the recency window’s benefit to the search restriction, not the scale estimator. Calibrating to 5% per look does not control false alarms across repeated looks, and autocorrelation degrades every method more than any noise-shape effect.