WHEN WORLD MODELS MISRANK: BUDGETED VERIFICATION FOR FINITE-SHOT QUANTUM CONTROL

A world-model planner can recommend a control that is worse than the incumbent. Detecting such errors with fresh system measurements appears attractive, but those measurements consume the same scarce interactions needed to learn the model. We study this trade-off in finite-shot quantum control using a replacement gate that compares a candidate and incumbent, charges both measurements to the interaction budget, and reuses their outcomes for learning. A frozen opportunity-cost experiment across four simulated tasks and 40 paired task–seed units compares this procedure with reallocating every verification shot to additional active learning probes. The gate improves equal-task normalized fidelity area under the budget curve by 0.1281 (95% CI [0.0990, 0.1564]) and final fidelity by 0.1797 ([0.1285, 0.2302]); both Holm-adjusted p-values are 2 × 10−5. Earlier held-out studies independently test unconditional replacement and alternative search baselines. A post-hoc diagnostic on 443 gated-trajectory decisions finds that measured advantage detects harmful candidates with AUROC 0.8212, compared with 0.5599 for predicted advantage and 0.4315 for pairwise ensemble disagreement; the ordering persists on a separate always-deploy trajectory set. A conservative Hoeffding gate provides a conditional family-wise replacement guarantee at a substantial acceptance cost. These results establish a regime in which spending interactions on deployment verification is more effective than using the same budget only for additional model learning, without claiming universal superiority of verification or general failure of uncertainty estimation.