From annotation to evaluation: what ChEMBL's potential_duplicate flag marks, and what applying it changes

Measurement independence is often inferred from document provenance. Across 100 randomly drawn ChEMBL_37 targets, 4,215 of 11,605 publication-publication same-compound-target pairs (36.3 %) met ChEMBL's published probable-repeat-citation criterion, abs(delta pChEMBL) < 0.02; 2,600 (22.4 %) were exact ties, versus at most 0.15 % in simulations of independent measurements under published replication error and three-significant-figure rounding. A preregistered blinded audit of 160 pairs found predominantly related documents. Its round-2 primary comparison missed its frozen significance criterion (p = 0.055), whereas the secondary reuse indicator met its criterion (OR 4.62, 95 % CI 1.225-19.73, p = 0.012). Two separately preregistered fixed-prediction benchmark analyses with matched random deletion met their frozen average practical-equivalence rules: CARA LO-GPCR, -0.003386 PCC; NTAB Chemprop, -0.002068 Pearson r. Effects were sparse and concentrated: 720 of 749 eligible NTAB strata contained no flagged test row, while the maximum-similarity bin carried 3,109 of 3,132 flagged test rows. The paper claims neither a new leakage type nor a new detector. The optional secondary NTAB retraining experiment was formally declined before execution for this release; no training-side result was inspected. The accompanying supplement contains the frozen confirmatory protocols, code, mappings, raw outputs, summaries, execution logs, technical audit, numerical checks and SHA-256 manifests. The pre-execution package was publicly timestamped before confirmatory execution at Zenodo DOI 10.5281/zenodo.21701846. OpenAI Codex and Anthropic Claude were used under human supervision as described in the manuscript; they are not authors or reviewers.