Spend Reviews Where Decisions Are Hard: Adaptive Review Allocation Under Fixed Conference Capacity

Large conferences increasingly face a mismatch between submission volume and expert reviewing capacity. Uniform policies respond by assigning nearly the same number of reviews to every paper, even though decision difficulty varies sharply across submissions. We formulate reviewing as fixed-budget threshold classification and study an adaptive allocation rule that sends each additional review to the paper with largest uncertainty index $I_i = v_i p_i(1-p_i)$. We establish two structural allocation results. First, when candidate papers have equal review counts, this index ranks papers exactly by the expected one-step reduction in decision uncertainty. Second, review demand scales with distance from the acceptance boundary: a paper at latent margin $\Delta_i$ requires order $\sigma^2/\Delta_i^2$ reviews for fixed error, and a continuous oracle pool obeys the exact budget-rescaling law $B_c^\star(\epsilon)=B_1^\star(\epsilon)/c^2$ when all margins are multiplied by $c$. Separately, we prove that predictable adaptive sampling preserves a Gaussian e-process and therefore supports a single terminal e-BH decision with finite-sample FDR control. In matched simulations with 2,000 submissions, the primary Adaptive-P method uses 80% of a four-reviews-per-paper baseline yet reduces total error from 4.32% to 1.92% at reviewer-noise CV 10%, while increasing power from 87.97% to 93.59%. A matched martingale/e-value allocation variant produces nearly the same operational performance, while its evidence-only e-BH layer is substantially more conservative. A dense label-preserving budget sweep further shows that pools with more mass near the decision boundary require more review capacity under high reviewer noise. The main implication is operational: conference review demand should be modeled as a function of decision difficulty, not submission count alone