Cross-Domain Benchmarking of Focus Measures for Smear-Microscopy Autofocus Under a Consensus-Audited Reference

Focus-measure recommendations for smear microscopy are typically established on one specimen type under one reference definition and one scoring rule, leaving their transferability untested. We benchmark 32 handcrafted focus measures on 26,100 z-stacks spanning five smear-microscopy domains at native acquisition dimensions under one frozen protocol. A fixed ten-voter plurality consensus (REF-B) supplies an identical reference target for every candidate, and two further constructions, single-voter exclusion (REF-A) and a non-derivative four-voter consensus (REF-C), quantify how far the recommendation depends on that target. Ten criteria covering consensus localization, curve shape, perturbation response and kernel cost form a declared cross-domain score; clustered bootstrap resampling, alternative aggregation, weight perturbation and controlled image resampling separate sampling variability from sensitivity to evaluation policy. Gradient responses lead under every analysis. Variance of Gradient ranks first under the primary policy in all 1000 clustered bootstrap replicates and in 67.9% of 1000 sampled weight configurations, while Brenner Gradient attains the lowest consensus deviation, 0.072 planes, at the lowest measured kernel cost. Localization ordering is preserved exactly when the reference is rebuilt from non-derivative voters alone (Spearman ρ = 1.000), although the reference plane itself shifts substantially. The benchmark gives a reproducible, auditable basis for selecting focus measures under stated reference, scoring and image-sampling conditions.