Fairness in histopathology AI — systematic review: extraction, audit, and verification artifacts (88-study corpus, manuscript v0.6.7)

Machine-readable evidence and analysis code for a PRISMA-2020 systematic review of fairness in histopathology AI. Contents. (i) The dual-coded, adjudicated extraction table for the 88-study analytical corpus (61 from the main search, 17 from the demographic-term update, 2 retrieved in October 2026, 2 suggested in peer review, 1 candidate resolved on full text and 5 from a re-screen of the keyword-filtered records; 89 records, one conference abstract linked to its journal paper; the earlier 62- and 79-study tables are kept alongside), with the pre-pass snapshots and the script that applies the October 2026 revisions, with the raw multi-coder outputs and inter-coder agreement statistics; (ii) the screening, de-duplication and post-hoc eligibility-audit logs; (iii) adjudicated QUADAS-2 risk-of-bias judgements for all 88 studies, with per-coder judgements for the original 62; (iv) the full consensus-endorsement coding pipeline — screened review pools, the pre-specified random sample (n=30, seed 20260913), the coding specification, two independent coder passes, disagreements and their adjudication, and tallies with Wilson 95% confidence intervals; (v) the 39-dataset demographic-metadata audit; and (vi) a claim-by-claim statistics register mapping every reported number to its source rows, together with the code that produces each value. Full texts of the reviewed papers are not included, for copyright reasons. Code is licensed MIT; data tables are licensed CC BY 4.0. A SHA-256 manifest (MANIFEST.sha256) covers every file in the package. This package is a working-tree snapshot of the manuscript as of 18 October 2026 (61 pp main + 45 pp supplementary). Manuscript v0.6.7 (18 October 2026). Audit of the four deep reviews against the current text (no numeric or citation error found; no coded value, count, grade or headline statistic changed). One sentence added to the Methods update-search paragraph pointing to the Supplementary Methods, which account for all 16 re-screen candidates; the Dawood 2026 reference carries its article pages and publisher dates (the publisher still assigns no volume or issue); audit record in notes/deepreview_audit_2026-10-18.md. v0.6.6 content, carried forward: Same manuscript text as v0.6.5; that version added data/context_fulltexts/, the full texts of eleven context-only works cited in the manuscript (final STARD-AI, CLAIM 2020 and 2024, TRIPOD+AI, PROBAST+AI, PRISMA 2020, QUADAS-2, Yang 2025, Bhattacharyya 2025, the EU AI Act and the 2021 GMLP principles) with a per-file licence list; this record's CC BY 4.0 licence covers the depositor's own materials only, and each of those texts keeps its rights holder's terms (three are publisher-copyrighted, one is non-commercial, one licence was not verified). v0.6.5 content, carried forward: Fourth deep-review response (every comment except the human audit, which has not been run); no coded value, count, grade or headline statistic changed. Table 12 caption and Table 10 footnote arithmetic corrected; PRISMA items 8, 13f, 20d and 23c completed; the Methods now separates the adapted QUADAS-2 appraisal from the full-text search for self-reported reporting-guideline adherence; the D6 attention metric is piecewise-normalized on [-1, 1] with an explicit dispersion formula; the final STARD-AI statement, CLAIM 2024, TRIPOD+AI, PROBAST+AI, PRISMA 2020, QUADAS-2, the EU AI Act and the 2021 GMLP principles were read in full as context (full texts are not deposited); ten references carry full author lists and one wrong co-author list was corrected; the 62 re-screen records without an abstract were reassessed (none added). v0.6.4 content, carried forward: One wording fix on top of v0.6.3 (the Confounding exemplars in the taxonomy figure, Results and Appendix S2 no longer cite a study that is coded institutional/selection rather than confounding); no coded value, count, grade or headline statistic changed. v0.6.3 content, carried forward: provenance correction, a record of a second LLM re-read, and a full-manuscript consistency sweep; no coded value, grade or headline statistic changed. This version also replaces v0.6.1 (the v0.6.1 supplement PDF printed an unresolved figure reference). The analytical corpus is 88 studies (89 records, 86 evaluate a model). Headlines: any fairness metric 25/86 (29.1%), external validation 47/88 (53.4%), overall risk of bias 7 low, 60 unclear, 21 high; 113 mitigation rows; feature-level harmonization is the only Strong family. data/coding_audit holds the pre-registered human-audit protocol, which has still not been run, together with the record of a second LLM blind re-read of that frame on 18 October 2026 (65 studies, 1170 worksheet cells, 113 mitigation rows; 81.6% and 80.3% agreement with the published coding; two of fifteen family grades sensitive; no published value changed), the author's approval of those outputs as a record (human-approved, AI-executed, not a human audit), and an empty human-verification ledger prepared for the human audit. The file with the sealed baseline values is withheld. A read of the whole manuscript against the evidence tables corrected eight stale statements in the prose (bias-category counts in the taxonomy figure and Appendix S2, a fairness-metric share, risk-of-bias shares and disagreement counts, a mitigation-study denominator, a TCGA count); the coding tables were unchanged. data/session_archive.zip archives the September 2026 writing sessions. The manuscript is 61 pages and the supplementary materials are 45 pages. Full texts of the 88 studies and of the other corpus papers are not included.