The calculation’s entry point

A model for concrete strength has a starting point before it has a prediction: the exact table fed into the calculation. Kim and Bajare’s audit of 419 papers locates a substantial gap at that point. In 211 studies, readers receive a citation to a data source without the complete dataset used. Selecting rows, cleaning values and merging tables can change the input. The audit’s most useful distinction exposes the distance between a bibliography entry and an executable starting point.[1]

Complete datasets were accessible for 90 studies and source code for 20. Reading the 90.7 percent without shared code as the proportion of wrong results would assign this audit a measurement it never made. The team did not run programs or recalculate predictions. It measured access to material needed to begin the same calculation. The design establishes an access bottleneck; ranking model correctness requires a further computational test.[1]

Testing the count itself

The classification procedure has its own check. A second assessor independently coded 84 randomly selected papers, with agreement of 96.4 percent for dataset categories and 97.6 percent for code categories. That supports consistency in how access was classified. Agreement between two assessors still leaves the usefulness of the files to an execution test. Reliability of the measuring procedure and correctness of an unmeasured result require different checks.[1]

The sampling boundary also carries the result. The search began in one index with a 2025 filter, retaining 38 entries later dated 2026 because they belonged to the initial retrieval. Image-based crack detection was excluded while tabulated mixture-variable studies were included. This design therefore cannot extend the 90.7 percent figure across engineering. Understanding the denominator directs the finding toward the research community actually sampled.[1]

From files to execution

For journal assessment, I would move the practical threshold from a sharing statement to versioned files. The authors’ recommendations for a persistent archive, the actual training table, source code and dependency information belong together. That deposit supplies concrete material for assessing whether an independent user can start the calculation. Teams working with compiled datasets face a particular obligation: identifying the original source leaves the transformations applied to it to be documented as well.[1]

This count cannot establish why access is missing. Legal restrictions, confidentiality or intellectual property could explain some unavailable files; the design does not test those explanations individually. The audit can therefore guide an access requirement without becoming a misconduct verdict. The unresolved question is whether an independent run can actually be completed using the available data and code. That test begins where this survey ends.[1]