NTR-Bench: Release-aware temporal variation and provenance limits in a natural-product-defined bioactivity benchmark

TR-Bench is the public reproducibility dataset and code package accompanying a release-aware analysis of temporal variation and provenance limits in a natural-product-defined bioactivity benchmark. It contains analysis-ready derived tables from ChEMBL-based historical and later-literature evidence, recorded-context comparability outputs, selection diagnostics, provenance-stress results, a strict-C1 repeated-publication evidence-accumulation analysis, a separate BindingDB cross-curation lane, compact Python reproduction scripts, source/version metadata, and licensing information. Raw third-party database snapshots are not redistributed. Version 1.4 synchronizes the public research object with the pre-submission EA1R1 context-integrity repair. Within the original 106-pair repeated-publication cohort, 21 non-C1 later documents were excluded across 11 pairs and same-document C1 context purification changed the effective sequence in additional retained documents. The corrected evidence-accumulation cohort contains 105 pairs; the descriptive fourth-publication cohort contains 63 pairs. The corrected random-order result remains compatible with order/estimator mechanics, so the scientific interpretation is unchanged. The main 769-pair temporal analysis, selection analysis, provenance perturbations, and BindingDB lane are unchanged. The portable workflow reproduces the manuscript's headline numerical results, including the 25,764-pair historical baseline, 769 recorded-context C1 pairs, median absolute temporal change of 0.509 pActivity, the corrected 105-pair evidence-accumulation analysis, and the 214-pair BindingDB cross-curation analysis.