Population Identity Preservation Prevents False Aggregation in Reanalysis of BACE1-Inhibitor Trials in Alzheimer's Disease

This preprint presents a reproducible failure-forensics study of public EPOCH and APECS verubecestat trial evidence. The study examines how population identity loss, over-compression, over-splitting, and context leakage can create false aggregation or block valid comparisons during clinical-trial reanalysis. The frozen methods branch reproduces six canonical biomarker relations, 4,739 dose-mapped adverse-event records, complete attribution of 741/741 dose-only multi-rate artifacts, exact closure of a 132-key filter-scope difference, restoration of four biomarker populations from 10 population-measurement rows, and exact biomarker and adverse-event population parity. The study does not claim biomarker-to-adverse-event causality, adverse-event causation of clinical failure, discovery of a previously unknown adverse effect, or safety or efficacy of any rescue candidate.