flatresaudit v1.0.0: reading an unchanged AUROC — audit code, retained outputs, verification package and Supporting Information for distance-based out-of-distribution detection

Code, retained outputs, protocols, verification package and Supporting Information accompanying the article “Reading an Unchanged AUROC: Invariance, Insensitivity and Infeasibility in Distance-Based Out-of-Distribution Detection” (prepared for submission to Expert Systems).

Contents: the flatresaudit Python package (three checks: ranking reach and realized reordering from image-level scores, separating low-reach responses from empirical protection at the tested dose; pre-data feasibility of corrected permutation inference; mean-shift projection, with construction-target status left to a separate provenance review); the full analysis pipeline (73 scripts) and every retained output; a verification script that checks 376 reported numbers; a map from every table and figure to its script, input, output and reproducibility level; the Supporting Information (proofs S1.1–S1.9, Tables S1–S10, LaTeX source and PDF); the internal protocols (none filed with a third-party registry) with a protocol registry of dates and SHA-256 hashes, the deviation and correction logs and a provenance record; split metadata, data-access terms and tested environments.

The audit diagnoses a detector’s scores and rankings; it does not certify clinical safety. The planted-ground-truth validation, including its confirmatory test and the comparison with confidence-interval and equivalence readings, is reproducible end-to-end from this archive (byte-identical outputs under the pinned versions). The dermatology study is verifiable from retained outputs and partially re-executable: images, the PanDerm checkpoint, embeddings, adapter weights, per-example scores, and two unpublished upstream modules are not included. Derived statistics and identifiers remain subject to the source datasets’ terms (DATA_LICENSE.md).

The revised main Results include the retained five- versus ten-seed calibration comparison. The multiway bootstrap intervals are exploratory and unadjusted. Adapter-external evaluation does not establish backbone-pretraining disjointness: image-, lesion-, and patient-level overlap with PanDerm pretraining has not been independently verified. The ISIC near-duplicate screen does not inspect the complete pretraining corpus.

MIT applies to author-written code and documentation. Third-party derived data remain subject to the original providers’ terms in DATA_LICENSE.md; the record license does not relicense third-party data.