Row-order misalignment in a public microarray deposit: an exactly correctable defect affecting 22+ re-analyses

Public expression databases underpin secondary analyses, resting on an implicit assumption: that the deposited file faithfully represents the data the authors analysed. We report a case that violates this assumption yet can be fully repaired. In the supplementary file of GEO series GSE47472 (Illumina HumanHT-12 V4; 14 abdominal aortic aneurysm neck specimens and 8 donor controls), the intensity block is displaced relative to the probe-identifier and detection-p-value blocks of the same file: the entire block is misaligned by 39,426 rows. Using eight donor arrays shared with a companion series (GSE57691) as anchors, we solved a unique bijective permutation pi from per-row 8-tuple keys (39,426/39,426 rows uniquely matched; maximum absolute difference after alignment = 0). The permutation was derived from the 8 control columns alone, yet after applying it the 14 AAA columns that took no part in the solution moved from a Spearman correlation of -0.008 to +0.990 between intensity and detection p-value. Probe-mean platform fingerprint against GSE57691 rose from -0.0086 to +0.9518. The deposited series matrix is an exact additive rescaling of the intensity block (slope 1.0000000, residual SD 2.67e-7) and therefore inherits the same misalignment. After correction, CD19 rose from -0.03 (P=0.93) to +1.65 (3.14-fold, P=0.0064), reproducing the original paper's qPCR result exactly, and cytokine-cytokine receptor interaction reached p.adj 7.7e-7; the uncorrected matrix yielded no KEGG pathway with p.adj<0.05. Routine quality control is entirely insensitive to this defect: the number of genes with FDR<0.05 changed only from 2,986 to 2,812. A literature audit (Europe PMC, 85 hits, 72 full texts parsed, 51 classifiable) found 22 papers that directly read the deposited matrix and 17 that merged it; as of 2026-10-23 no published work questions this dataset. We provide the permutation, the repaired matrix, and diagnostic statistics, and recommend a within-file cross-block consistency check at both deposition and re-use. This is a single case; the generative mechanism of the permutation cannot be attributed to any public platform field, and we do not speculate. We do not allege fabrication in any form: the observed defect is a numerical inconsistency within and between public files, and a repairable one.