Abstract
<jats:p>Absorption, distribution, metabolism, excretion and toxicity (ADMET) measurements are often reported only as upper or lower bounds because an assay does not resolve activity beyond its tested concentration range. These inequalitylabelled observations contain useful directional information, yet dropping them loses signal and threshold substitution changes the regression target. Interval likelihoods retain the information but provide no direct control over the extent to which censored observations influence exactvalue fitting. We introduce CensorADMET, which minimises an exactlabel objective with an exactonly prediction anchor while targeting a prespecified aggregate probabilitydeficit budget for censored intervals. Across ChEMBL 36 inhibition cohorts evaluated under chemical and assayprovenance shifts, the primary anchored satisficing setting reduced onesided bound violations by 0.15–0.29 with a median exactlabel MAE cost of at most 0.04 punits. Plain Tobit interval likelihood achieved larger violation reduc tions but larger MAE costs; concentrationspace AFT provided a competitive fixed operating point but no tunable deficit target. The same tradeoff was observed under simultaneous assayand scaffold shift and was directionally consistent in an exploratory cohort of solubility, clearance, and halflife measurements. CensorADMET therefore allows practitioners to specify and audit a modelimplied censorconsistency target rather than accept the operating point induced by an unweighted interval likelihood or indirectly selected loss weight. Scientific Contribution. CensorADMET introduces a constrained censoredregression objective whose userspecified aggregate probabilitydeficit budget parametrises a targeted accuracy–con sistency operating point. It evaluates this budgettargeted control across natural censoring and chemical and assayprovenance shifts, separating bound consistency from exactvalue accuracy, uncertainty calibration, and hiddenvalue recovery.</jats:p>