Structure over sound: A free box count out-predicts in-hive acoustics for honey-bee colony strength

In-hive microphones are widely promoted for non-invasive estimation of honey-bee colony strength (frames of bees), and benchmarks report near-perfect accuracy. Whether that accuracy survives on unseen colonies and apiaries decides deployability. We re-analysed two open multi-sensor datasets, UrBAN (ten urban colonies) and MSPB (53 rural colonies at two apiaries), harmonised to a shared 20-feature acoustic space with in-hive temperature and humidity. Random forests and hierarchical mixed models were evaluated under progressively stricter validation, from random k -fold to leave-one-apiary-out and cross-dataset transfer, with the acoustic, environmental and structural contributions separated by ablation. Most of the reported skill is not acoustic. Under leave-one-apiary-out validation the acoustic features alone reach R 2 = 0.17 , in-hive temperature and humidity alone 0.56, and the beekeeper’s box count alone 0.71; the box count from the previous visit, fixed before the label existed, still reaches 0.42 alone and 0.55 with climate. Adding acoustics never improved accuracy materially under any configuration tested ( Δ R 2 between −0.12 and ＋0.02). The box count’s 90% conformal intervals keep near-nominal empirical coverage across apiaries, while those of the sensor model cover 70% of cases in one transfer direction and 99% in the other. In an exploratory cross-dataset comparison, limited by a 17-assessment audio-matched UrBAN sample, the box count transferred in both directions and the acoustic features in neither. A plausible reason is sensing geometry: the microphone sits in the bottom brood box, which barely varies, while almost all population variation is above it. The box count mixes physical capacity with management decisions. It is therefore a diagnostic baseline and not a measurement instrument, and the finding is that a single bottom-box microphone adds nothing to it, though acoustics retain value for other targets. The conclusion rests on the within-MSPB cross-apiary comparison and is limited to a single bottom-box microphone, the shared feature space and the frames-of-bees target.