Defining the Data set Defines the QSAR Claim

Machine learning has greatly expanded QSAR modeling, but predictive claims still depend on choices that are rarely documented: how chemicals are represented, how end points are defined, and how evaluations are designed. In the era of benchmarks and foundation models, inconsistent standardization, unclear rules for combining measurements, and hidden information leakage routinely inflate reported performance while obscuring weaknesses that matter for real-world applications. We propose data set contracts: executable, auditable documents that explicitly declare chemical processing rules, end point definitions, aggregation logic, data splits, and leakage diagnostics for the intended prediction scenario. These contracts are feasible with current open-source tools and would shift the field from architecture-centric comparisons toward claims that are transparent, reproducible, and trustworthy.