Artifact for "A Ceiling-Aware Complete Factorial Study of ImageNet Pretraining and Backbone Fine-Tuning in Compact Ultrasound Classifiers"

Analysis plan, training and analysis code, per-run integrity assertion records, per-seed split fingerprints, and the canonical statistics report for a 120-run prospectively specified confirmation of a complete pretraining-by-backbone-fine-tuning factorial in compact ultrasound classifiers. The batch is 3 plain CNN backbones (MobileNetV3-Small, ShuffleNetV2-x1.0, ResNet18) x 4 factorial cells (pretrained/random initialization x fine-tuned/frozen backbone) x 10 seeds, run in a single environment on 2026-08-19 with no technical retries. Every run directory carries the dataset manifest, the split cache, per-epoch metrics, an environment record, and a set of machine-checked assertions covering the split, the graph surgery, the optimizer partition, and the frozen-parameter guarantees. verify_archive.py re-derives every recorded SHA-256 from the archived manifests using only the standard library, so anyone can confirm the copy is intact before trusting the analysis. plain_primary_analysis.py reproduces the canonical statistics report exactly, including the BCa bootstrap intervals, because the bootstrap RNG seed is fixed. The ultrasound images themselves are not redistributed. Dataset 1 aggregates three public collections; each run directory records the file ordering and labels used, so the splits can be reconstructed once the images are obtained from their original sources under their own licences. Licensing: the run records are released under CC BY 4.0; the code and documentation are released under the MIT licence (see LICENSE in the archive).