Molecular predictions increasingly influence which compounds are screened, synthesized, optimized, or advanced, yet the evidential conditions surrounding those predictions are often separated from the outputs that enter experimental workflows. A probability, rank, or predicted property can therefore travel farther than its calibration evidence, applicability limits, data provenance, model version, and intended use. This article develops a proposed molecular prediction evidence passport as an original prediction-governance construct for keeping those elements attached to a bounded prediction event. The approach synthesizes methodological insights from molecular machine learning, uncertainty quantification, biomedical data stewardship, reporting guidance, and pharmaceutical decision practice. The passport is defined as a versioned, machine-readable and human-interpretable record that identifies the prediction object; preserves model, code, representation, and data lineage; distinguishes model confidence from calibrated uncertainty; records applicability conditions and unresolved uncertainty sources; states the experimental decision the prediction may inform; and links subsequent experimental consequences back to the originating record. The central contribution is not another performance metric, but an architecture for preventing benchmark performance, explanation, plausibility, or organizational approval from being misread as experimental confirmation, mechanism, therapeutic value, or deployment readiness. The construct may support more inspectable handoffs across screening, optimization, and candidate selection, while enabling later audit of what was known, assumed, changed, and acted upon. Its value remains conditional on schema design, semantic interoperability, data quality, human oversight, and prospective evaluation. The passport is therefore presented as a testable governance proposal rather than a validated standard, regulatory artifact, or autonomous decision system.