Additional file 5. Robustness analyses and source data

Contents
1. Additional_file_5_robustness_summary.pdf
   Human-readable summary of the fully out-of-fold representation analysis,
   annotation-training ablation, duplicate-conservative eligibility analysis,
   group out-of-fold validation, source-batch holdout, deployable-variable
   analysis, ZhengHou paired comparisons and complete-case selection audit.

2. Additional_file_5_robustness_source_data.xlsx
   Aggregate machine-readable source tables for the analyses above.

3. same_split_image_backbone_benchmark/
   Aggregate locked-test performance, paired full-minus-image differences,
   figure source data and execution metadata for the direct comparison with
   ConvNeXt-Tiny, ResNet50, EfficientNet-B0, DenseNet121 and Swin-T image-only
   models. All direct comparisons use the same 356-patient cohort, 1,053-image
   manifest and locked 51-patient test set.

Privacy and scope
- No patient identifiers, names, filenames, storage paths or raw images.
- No patient-level predictions or embedding arrays.
- These analyses were not used to select a new headline model; the image-
  backbone benchmark is exploratory and its paired intervals are descriptive.
- Source-batch holdout is internal transportability analysis, not external validation.
