OrderGap: statistical analysis plan v1.0 (frozen 27 September 2026) and deviations log

Statistical analysis plan for OrderGap, a study of models that predict resistant Enterobacterales (resistance to ceftriaxone, ceftazidime, cefepime or piperacillin-tazobactam). It compares models trained on positive cultures with models trained on every ordered culture, evaluated when the culture is ordered. Data are the four public ARMD releases (Stanford Health Care, Mass General Brigham, UT Southwestern Medical Center, ECU Health) and MIMIC-IV 3.1 (Beth Israel Deaconess Medical Center) as external validation. The plan calls the study "Paper A", its working name. The plan was frozen on 27 September 2026, before any model was trained or any test partition was read. Freezing was recorded as an annotated git tag holding the plan's SHA-256, and every training and evaluation run recorded the SHA-256 of the plan in force. This record was deposited on 30 October 2026, after the analyses ran. It contains the frozen plan and the three later versions of its deviations log, each matching the digest its runs recorded. README.md lists the digests and the runs that pinned them, and discloses one test read that was not logged at the time. No file contains patient-level data.