Answers are grounded in this paper and cite their sources.
Preserving a monitoring decision after missing‐data repair does not establish that the decision is correct. We develop an audit for a fixed prediction rule that combines split conformal stability checks with paired error attribution and action‐specific failure accounting. Generated experiments assess calibration under exchangeability; historical hard‐drive cohorts assess the resulting actions. In the outcome‐ascertained 2014 cohort evaluated with a frozen 2013 pipeline, 42 operator‐recorded failures occur among 31,384 normal actions: 0.134% of normal actions but 55.26% of observed failures. Repair corrects 46 reference errors and introduces 28, while 108 errors remain unchanged. A further 2015 test enrols 18,595 devices absent from the retained study identity union and locks baseline predictions and comparator identities before archived follow‐up is parsed. All 19 observed failures receive normal actions, with no reference‐to‐repaired decision flips. The normal‐action failure fraction among ascertained outcomes is 0.1078%, whereas observed failure miss is 100%. Accounting for 32 unresolved outcomes gives a failure‐miss identification range of 73.08%–100%. In both temporal tests, the stability gate adds no selection beyond the label gate, and confidence rankings reproduce the combined selection. The audit identifies inherited errors and redundant components without demonstrating a ranking advantage. The evidence concerns historical transport under imposed input loss; sparse native missingness and rare calibration failures limit claims about field reliability.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: