Provenance record for "The Data Channel, Not the Writing Channel"

Evidence record supporting the case study "The Data Channel, Not the Writing Channel: Where AI-Introduced Errors Entered a Scientific Manuscript" (M. Kozlowski, 2026). An adversarial AI audit of a radiology-AI preprint asserted, in July 2026, that the manuscript had misnamed two of the five models it evaluated, on the stated ground that one of the names did not exist at the provider. The author accepted the correction. The manuscript had been right and the audit was wrong. This record contains the material needed to verify that independently. Four lines of evidence are provided. First, code extracts showing that the batch runner invokes its adapters without a model argument, so the adapter defaults executed, and that those defaults name the models the manuscript originally reported. Second, filesystem timestamps with SHA-256 checksums establishing that both adapters were last written before the first generated output appeared and were never touched after the run finished, which excludes editing after the fact. Third, the status manifest for all 300 inference calls: sixty successes for each of the two disputed models establish that the identifiers were valid. Fourth, provider records confirming that both models exist and that the release date the audit removed as fabricated is correct. The record also contains the audited manuscript across five states, showing the model names correct in April, incorrect after the July audit, and restored after forensic re-examination; both audit reports in full and unedited, including the passages that were wrong; the released analysis data against which the manuscript's tables were recomputed; and an adversarial audit of the case study itself, which found one serious error in it. The two audit reports are in Polish and are reproduced verbatim rather than translated, because the case study makes claims about what they assert. See audit_reports/NOTE_ON_LANGUAGE.md, which lists the relevant passages with translations. All other documentation is in English. Clinical source documents are not redistributed. MIMIC-CXR images remain available from PhysioNet under credentialed access. Companion record: the Study 2 reproducibility bundle, DOI 10.5281/zenodo.21427298.