Abstract
Structured decision models that return typed outputs, such as choices, judgments, and scores, enable efficient decisions but often leave the evidence behind each call unauditable. We study this gap in active feature acquisition (AFA) under hard budgets, where each feature incurs a cost and a prediction must be made after acquiring at most B features. We present JevSpec, a training-free framework that makes the acquisition policy itself an auditable typed decision specification, with a per-instance receipt. JevSpec (i) unifies tabular data and a natural-language goal into an auditable feature table; (ii) defines three information-gain strategies—static, conditional, and discriminative expected log-likelihood gain (ELLG)—plus random and sequential baselines under one shared predictor per dataset; and (iii) types the resulting decision points with Choice, Noul, and Score primitives. Acquisition does not call a learned decision model; the exported specification is what a reviewer can replay. Under a hard-budget protocol, static information gain reaches 0.806 accuracy at B=5 on MiniBooNE, above random (0.722) and sequential (0.735). On Diabetes, information-gain strategies reach 0.790-0.798, against 0.702 for random. Discriminative ELLG is sensitive to predictor capacity and collapses toward random under a lightweight linear backbone.