Confidence Is Not Enough: Decision-Value Finding Selection Beats Confidence-Gated Test-Time Compute

Test-time compute methods decide how much extra computation to spend by reading the model's own confidence. We show that signal is invalid exactly when the model lacks the information it needs to answer: on interactive MedQA-USMLE the confidence–correctness correlation is near zero under minimal information and becomes predictive only as the record is read. 16$\times$ self-consistency at low information buys nothing ($+0.017$, $p=0.72$) while revealing two findings gains $+0.197$ ($p<2\times 10^{-4}$), even beating a gold-aware compute allocator ($+0.092$, $p=0.044$). We make finding selection first-class and control it by the classical decision value of information: the finding most likely to flip the model's greedy answer. The criterion is training-free, gold-free, and survives miscalibration because a flip is a behavioral event. The prevailing selection criterion, greedy entropy reduction (the literature's “EIG”), is empirically indistinguishable from confidence gain and inherits the same pathology. Decision-value and entropy-reduction coincide under calibration and diverge under insufficiency, at two layers—selection and stopping—with which layer pays off set by how the belief is miscalibrated: selection wins under rank miscalibration ($+0.13$–$0.16$), stopping under level miscalibration, where the continuous signal saturates ($72$–$95\%$ of states $<10^{-3}$) but the binary flip transfers ($+0.05$–$+0.13$ across Qwen2.5 scales, $+0.24$/$+0.29$ on reasoning models, at matched reading budget). One decision-value currency arbitrating read/compute/stop Pareto-dominates the deployed controller ($+0.255$, $p<2\times 10^{-4}$) and clears the entire pure-reading frontier, beating its best point by $+0.041$ at $45\%$ less budget. Scope: four-way multiple-choice reading over an available record, against a self-consistency compute baseline; we do not evaluate acquisition of genuinely unobserved outcomes, nor compute controllers built on verifiers, process rewards, or search.