Treatment phase and sample retention constrain external validation of drought microbiome scores

Preprint; not yet peer reviewed. Drought treatment history and current water limitation define different prediction targets. We evaluated whether a classifier trained in one published study retained an elevated score after rewatering in an entirely held-out rice experiment, and audited metadata and sample retention relevant to validation. We reconstructed 424 rice time-course profiles, including 157 post-rewatering profiles carrying a drought-history label. The primary random forest, trained on grass rhizosphere profiles, had a mean visit-specific active-drought area under the receiver operating characteristic curve of 0.844. Its baseline-adjusted post-rewatering score difference was 0.009, with a conditional tub-bootstrap 95% interval of −0.032 to 0.051, based on four prolonged-drought and eight control tubs. The endosphere adjusted endpoint was unavailable because baseline visits were incompletely represented. Only 15 of 80 original endosphere baseline profiles were retained; these were exactly the 15 meeting a 10,000-count threshold under the original processing. In a separate sorghum audit, 60 of 521 classifiable profiles differed between binary treatment labels and reconstructed irrigation phase. These results identify limits of the validation target and sampling support, without establishing absence of a microbial drought legacy or a new biological mechanism. Files: MANUSCRIPT.pdf is the preprint; Supplementary_File_1.zip contains the reproducibility materials; SUPPLEMENTARY_METHODS.md provides Supplementary File 2. No funding was received for this research. The author declares no financial or other competing interests. Assistance and original-data attribution are documented in the manuscript. Licence scope: CC BY 4.0 applies only to new author-held manuscript text, original plots, analysis code and derived analyses. Third-party source records retain their original terms; no additional rights are granted in third-party materials. Original processed-count archives are referenced by pinned source URLs rather than redistributed.