Decision Reliability When the Estimate May Target the Wrong Quantity

Abstract

Threshold decisions are commonly accompanied by uncertainty about an estimator while leaving a more basic question unstated: whether the estimator targets the quantity the decision actually concerns. This paper separates three sources of decision unreliability. Sampling uncertainty belongs to a conditional sampling law and is reduced by better measurement. Specification disagreement belongs to a governed set of admissible constructions and is addressed through agreement, set restriction, or robust reporting. Target displacement is the unknown difference between a construction's estimand and the decision-relevant target; it is an identification problem rather than an additional sampling variance. The framework distinguishes conditional clearance probabilities, governed predictive probabilities, and weight-free adversarial support. A two-point counterexample shows that predictive distributions with equal means and variances can have threshold-clearance probabilities of 0.5000 and 0.1004, so moment-matched Gaussian reporting is not generally licensed. Preserving the distinctions instead, we derive exact predictive probabilities, verdict-specific displacement tolerances, vanishing-uncertainty limits, and remedy non-equivalence, and state an information boundary specifying the strongest output supported by the available information. A consumer-lending application shows that specification, precision, population alignment, threshold geometry, and target displacement act on different primitives and cannot substitute for one another. The contribution is not a new reliability index, but an architecture separating decision failure modes by epistemic status, governance requirement, licensed output, and remedy. Data availability. The consumer-lending illustration uses the e-Car data set published by Stanford University Press as companion material to Robert L. Phillips, Pricing Credit Products (2018), and available without charge from the publisher. The workbook is not redistributed here; the manuscript's appendix reports its SHA-256 checksum, sheet, and row and column counts so that any reader who downloads it can confirm they hold the identical file. All analysis scripts and their derived outputs are available from the author.