# Aggregate extraction and screening provenance

All numerical data are published aggregate trial facts. No patient-level records, private data, or source article copies are included. The annotations and code are original. Bibliographic metadata identify sources for verification. The data license covers these annotations and compiled factual observations; it does not relicense the cited articles.

`trials.csv` records the seven eligible sepsis-specific alert trials and the excluded borderline continuous-monitoring trial from the earlier review. Its `source` and `notes` columns identify extraction locations and access limits. `primary_individual` means eligible for the individual-randomization sensitivity analyses; cluster trials remain in the evidence map. Hooper mortality and Kijpaisalratana mortality numerators were extracted from the 2025 review's forest plot, not verified in the original full tables. All figures entering the learned individual-trial synthesis were verified in original trial reports. Hooper's endpoint is mortality at 28 days or discharge according to the Cochrane extraction; Downing reports in-hospital mortality at 30 days; NAVOY reports death during ICU stay. These horizons are not interchangeable fixed-time all-cause outcomes.

`reported_estimates.csv` records SCREEN's reported adjusted results. Code copies those reported estimates into its output; it does not independently re-estimate them. Their trial-level patient data are not in this bundle. Raw-count comparisons are demonstrations of why a cluster/time-adjusted result cannot be reconstructed from a two-by-two table. The alert-only population is selected after assignment and is not a substitute for the all-eligible population.

`update_screening.csv` records every title screened from the broad Europe PMC search whose first publication date is June 1, 2024 or later, through November 3, 2026. One investigator screened them; there was no duplicate independent screening. Selected borderline records were read at abstract level, with original full methods where available. Most title exclusions plainly concern drugs, biomarkers, retrospective prediction, basic science, reviews, pediatric care, or unrelated interventions. The generic title-exclusion reason means exactly that; it does not imply full-text review of those records. The source review supplies pre-cutoff trial coverage; this is an update and reanalysis, not an independent multi-database review from inception. `search.json` supplies the exact search and page URLs. Only metadata and original decisions are redistributed, not abstracts. Search access date was 2026-11-04 UTC.

The model-class grouping is an explicit operational definition: the trial's fitted multivariable prediction system is learned, including Epic's logistic model, whereas direct clinical threshold criteria such as SIRS and qSOFA are rules. This distinguishes deployed mechanisms, not whether a clinical score originally used data in development, whether a model adapts during deployment, or whether its training methods and performance are shared. A sensitivity excludes the logistic model. The trials compare whole alert/response pathways, which can include telephone notification, pharmacists, training, and feedback; across-trial subgroup differences cannot estimate a causal effect of algorithm class.

The code uses no random draws. No deaths or missing values are imputed. NAVOY's reported ICU deaths are retained; the three patients still hospitalized at trial closure are not reclassified. The zero-event Downey cell receives a 0.5 correction in the explicitly invalid naive-count demonstration only. The main analysis excludes cluster trials from independent-patient meta-analysis and does not replace unavailable cluster-adjusted estimates with naive estimates.

`screen_periods.csv` transcribes aggregate numerical facts from Arabi et al., Supplement 2 eTable 16, PDF page 29, visually checked against the original publisher PDF. Zero denominators encode arms absent in a period; they are not invented patients. The period-seven total-control cell has malformed punctuation in the source, but the COVID/non-COVID components and totals agree with 53 deaths in 2,028 controls. `code/period_audit.py` reconstructs all totals and computes descriptive point contrasts only. Periods without both arms are excluded from stratified contrasts. The original supplement itself is not redistributed.

The automated integrity check flags the first digits of `record_index` in `update_screening.csv`. These are bounded row positions in a search result, not measured quantities expected to follow Benford's law. The zero index has no leading nonzero digit. Their nonconformance is a property of the indexing and date selection, not a scientific anomaly.
