IP102 Image-Family Screening and Evaluation Independence Audit — Anonymous Review Artifact

This repository provides the reproducibility artifacts for an anonymous research study auditing evaluation independence in the IP102 insect pest recognition benchmark. The study investigates potential overlap between training and test images and examines how it affects interpretations of reported classification accuracy. Using perceptual hashing and image-similarity screening, the audit identifies 19,673 candidate cross-split image pairs and flags 4,099 test images (18.1%) as potentially sharing image content with training observations. A stratified manual review of 300 candidate pairs found 297 visual matches. Accounting for the candidate-population weights of the 20 sampling strata, the estimated visual-match precision was 98.99% (95% stratified-bootstrap interval: 97.59%–100.00%). The screening results are also used to examine how excluding flagged test images changes evaluation outcomes for ResNet18 and EfficientNet-B0 classifiers. Repository contents include: Image-level identifiers and audit manifests for the IP102 classification partitions. Candidate cross-split image matches and associated similarity information. Candidate image-family and flagged-test-image identifiers. Stratified manual-review records and sampling information. Supporting validation materials to reproduce and check the audit. The repository does not redistribute the original IP102 photographs. Researchers should obtain the images from the original dataset source and use the provided identifiers to reproduce the analysis. The reported screening results raise concerns about potential image-level overlap and evaluation sensitivity. They should not be interpreted as a causal estimate of accuracy inflation or as evidence of identical biological specimens in every flagged pair. This archive is provided as an anonymous review-stage research artifact supporting reproducibility and independent verification. Author attribution and publication information will be updated following peer review.