Adaptive feature acquisition under diagnostic misspecification

Active feature acquisition selects which costly observation to obtain next using information already observed. We examine a standard greedy conditional-mutual-information policy in a fully specified finite Bayesian model. Four feature channels provide target signals and noisy regime hints; the primary comparison uses exactly two feature calls and an optimized context-conditioned static pair. Five constructed worlds expose complementarity, equal independent channels, redundancy, shared error and a deliberately misspecified diagnostic. The matrices, seed, sample sizes, source hashes and decision table were recorded in a local freeze receipt before the archived run, according to the project’s provenance record. Across 50,000 generated instances and 500,000 paired method evaluations, the adaptive policy improved log loss by 0.03649 nats per trial in the complementary world, with a 99 percent paired fixed-block bootstrap interval of 0.02813 to 0.04439. Accuracy increased by 2.87 percentage points. Equal-independent and redundant controls gave exact policy identity. The shared-error world retained a smaller positive contrast of 0.01558 nats, while a shift in diagnostic semantics reversed the effect to minus 0.19671 nats and reduced accuracy by 20.41 percentage points. A subsequent audit independently reproduced all five exact expected contrasts, verified the complete archived rows and recomputed all 100,000 primary posteriors and acquisition sequences. These results provide a transparent implementation and failure-case benchmark for an established acquisition rule. They do not establish a new perspective mechanism, optimal adaptive planning, total-compute superiority or transfer to trained models or natural cognition. The adverse result is part of the finding, not an exception removed from the report.