IP102-TailDet: A budget-aware tail completion benchmark for long-tailed insect pest detection

The IP102 benchmark is the standard testbed for insect pest detection, but its detection split is smaller than its classification corpus by roughly a factor of four and is severely long-tailed, and we show that the perceptual-hash de-duplication routinely applied to it is, if left unconstrained, actively harmful: because most IP102 images share a single-specimen composition, loose hashing merges images across distinct species and disproportionately prunes the tail classes that most need data. We audit IP102 for near-duplicate cross-split leakage, class-coverage deficit, and watermark contamination, and use the audit to build IP102-Clean, a leakage-controlled base with class-constrained de-duplication that recovers 1043 wrongly discarded images while eliminating exact cross-split leakage. The audit comprises these three signals only; no automated label-noise or box-quality signal was run for this release, and none gates the pipeline. On this base we introduce a budget-aware tail-completion procedure, IP102-TailDet, that mines human-verified bounding boxes for under-represented species from the much larger IP102 classification corpus under a fixed admitted-image budget, allocating that budget by a deficit-proportional per-class quota capped by eligible corpus supply, decoupling localization (from a detector proposal) from species identification (from the classification-corpus label) and keeping every review human-gated. Completion increases the number of classes represented by at least 100 detection images from 44 to 63 of 97 and shrinks the rare tier (<50 images) from 30 classes to a single species, without touching the original validation split. We release a construction protocol specifying how the evaluated IP102-Clean and IP102-TailDet data conditions were built from the public IP102 source, and report directionally consistent completion gains across off-the-shelf YOLOv8, YOLO26, and RT-DETR detectors under the present IP102 protocol. We repeat the Clean base versus completed comparison over three training seeds on the two YOLO families to show that the gain is not an artifact of a single seed.