# Decisions and deviations from the dated research plan The plan is not a registered protocol. The investigator had read the earlier review and SCREEN abstract before writing it. The final work updates that review's trial frame rather than claiming an independently exhaustive search of all databases from inception. All 694 post-cutoff titles in the retrieved records were screened. The full 2,725-record search is retained privately for provenance; its exact query and page URLs are public here, while the public table includes the update records only. The stricter eligibility rule in the plan excludes the continuous generic vital-sign-monitoring pilot by Downey from the sepsis-specific evidence map; it remains in a broad-frame sensitivity. CONCERN is excluded because its alert predicts general deterioration, despite its important randomized mortality finding. PRONTO compares procalcitonin with usual NEWS2-based care, not an automated alert with no alert. APT-Sepsis uses a broader maternal prevention and care program with a colour-coded observation chart. SEPTIC, MedGo, prehospital sepsis and PRECISE implementation/protocol reports do not supply eligible outcome contrasts. These are scope boundaries, not findings that those interventions fail. Full methods and a cluster-adjusted mortality estimate were unavailable for Kijpaisalratana. The plan prefers adjusted cluster effects and permits no pooled headline when estimands are incompatible. Accordingly SCREEN and Kijpaisalratana are not entered into an independent-patient efficacy meta-analysis. SCREEN's adjusted RR, raw-count RR and alert-subset RR are presented separately. No overall cluster-adjusted pooled causal estimate is claimed. The three individually randomized learned-model trials are summarized conditionally using the earlier review's mortality extraction choices: hospital deaths in Shimabukuro and Tarabichi, ICU deaths in NAVOY. The mixed horizon is a limitation, and this subset synthesis is exploratory, not a prospectively specified confirmatory test. A preliminary scratch calculation on these three counts preceded final code construction and gave OR about 0.65 with a wide t interval. Final analytical choices and every subsequent analysis are disclosed. A hospital-only sensitivity and an alternative Tarabichi 28-day endpoint are included. No outcome was selected because it strengthened evidence for or against benefit. The final implementation uses modified Knapp-Hartung intervals, with variance scale max(1,q), plus normal REML, fixed-effect normal and DerSimonian-Laird normal sensitivities. The plan named Hartung-Knapp and alternative heterogeneity estimation but did not specify that modification. Its purpose is to prevent an anomalously narrower t interval when q is below one. Few-study t intervals can be extremely wide; the hospital-only normal result supports benefit and is reported explicitly. Leave-one-out, RR instead of OR, horizon and model-class sensitivities are all retained. No formal publication-bias or meta-regression test is run with these few studies. The deliberately invalid naive-count analyses including cluster trials are methodological demonstrations and a numerical comparison with the earlier published calculation. They are labelled invalid in result keys and cannot support an efficacy claim. They do not re-estimate the published adjusted SCREEN effect, nor establish whether the primary adjustment is correct. Without individual or cluster-period data, that question remains open. After the main calculations, the original SCREEN supplement became accessible. We added an explicitly exploratory arithmetic audit of its period/COVID aggregate cells, calculating crude ratios in the full and concurrent-period samples and descriptive Mantel–Haenszel point ratios. This was not in the original plan. The analysis checks a numerical implication of aggregation, not clinical efficacy or valid uncertainty; it supplies no patient-independent interval and does not treat post-assignment COVID classification as a validated baseline confounder.