How reliable are Boltz-2 fragment co-folding rankings? Scrambled-control variability and model sampling in a RUNX2 case study

Data deposit for the manuscript "How reliable are Boltz-2 fragment co-folding rankings?Scrambled-control variability and model sampling in a RUNX2 case study". The archive holds the Boltz-2 predictions the manuscript reports, the input files thatproduced them, the analysis data every reported number is read from, the tables as theyappear in the article and its supplementary material, and the scripts the article names. It contains 512 prediction run directories: the 479-run tree the study analyses(enumerated in Supplementary Listing S1 and classified in census_output.txt), 11 RUNX2331-360 boundary-extension runs, 11 runs of a TLE1 receptor comparison, and 11 runs of afurther control series. Every run directory carries the parameter record written for thejob; 510 of the 512 also carry the per-model confidence JSON and the predictedcoordinates, and the two that returned no model output are named in census_output.txt. Contents: README.md (a description of every file), LICENSE-DATA, census_output.txt,predictions.zip, inputs.zip (677 input YAML files), data.zip (14 derived datafiles and eight deposited reference structures), tables.zip (17 CSV files, one pertable of the article and supplementary material, each carrying its caption),code.zip and provenance.zip (the four analysis and job scripts the article names). Original prediction seeds were not retained, so reproducibility refers to the reportednumerical analyses rather than regeneration of the original stochastic trajectories.