Verification-Conditioned Abstention: Withholding Unverifiable Claims in Labor-Market Decision Support

Abstract

Selective prediction, learning to defer, and conformal methods all condition the decision to abstain on the model’s estimate of its own correctness. We define and deploy an alternative. In verification-conditioned abstention the selection function is a decidable predicate over an external authoritative registry: a claim is emitted only if a specific public record entails it. The selector is predictor-independent, externally re-executable by the affected person, and fails closed on data absence rather than on a tunable threshold, so coverage is not a dial the operator controls. We instantiate this for United States H-1B sponsorship claims, grounded in two first-party federal registries, and show that the resulting coverage is a functional of the registry rather than of the model. Across 164,070 employers and 3,871,927 certified labor condition applications, 74.94% of employers admit a corroborating record (95% CI [74.73%, 75.15%]). Availability is strongly structured: it ranges from 58.9% to 95.8% by employer filing volume and from 41.3% to 79.3% across occupational groups. Because the selector does not read the model, this disparity is not addressable by any modelling improvement. One finding inverts the expected direction. Staffing and consulting intermediaries are more legible to the predicate than direct employers (86.8% versus 73.5%), so a verification-grounded system speaks most confidently about the employers a worker has most reason to treat with caution.