Pharmaceutical knowledge bases frequently join structural records, substance descriptions and bioactivity observations through molecular identifiers whose normalization rules are rarely identical. A canonical structure can therefore provide reliable computational matching while obscuring distinctions relevant to stereochemistry, tautomerism, salt composition and the material actually tested. This conceptual and methodological analysis examines published molecular-curation procedures, chemical databases and predictive-data studies, and reconceptualizes duplicate detection as an endpoint-conditioned relation between provenance-bearing records. It distinguishes graph representation, assigned chemical state, material composition, observed assay entity and prediction unit, and proposes auditable rules for preserving these distinctions during integration. Competing explanations include assay heterogeneity, identifier errors and data-partition leakage. The resulting decision specification is offered for prospective expert adjudication and controlled modeling comparisons; its accuracy and utility have not been empirically validated. Chemical identity must be established at the level required by the intended pharmaceutical inference before molecular records are merged or prediction labels pooled.
IJPRAS
