Experimental reference version sensitivity of fixed SAMPL8 pKa predictions in a selected remeasurement subset

This research compendium accompanies the Research Note “Experimental reference version sensitivity of fixed SAMPL8 pKa predictions in a selected remeasurement subset” by Sho Miyazaki. It contains the manuscript and Supplementary Information, preserved upstream sources, frozen analysis inputs, Python scripts, machine-readable results, figures and tables, and provenance and checksum records. Experimental reference values form part of the provenance of a computational benchmark. We examined how replacing the experimental-reference version affects fixed SAMPL8 pKa predictions for three compounds with one-to-one post-challenge SiriusT3 remeasurements. Selected state transition (SST) assignments and predictions were held fixed for S8-2, S8-5, and S8-22 and four complete-case calculations: COSMO-RS one pKa fit, EC-RISM, RFE-uESE-extra, and Chemaxon. Replacing the original references with SiriusT3 values changed all 12 cell-level absolute errors and the four three-compound mean absolute errors (MAEs). The subset MAE ordering was not fully preserved: COSMO-RS one pKa fit and Chemaxon reversed order, whereas EC-RISM and RFE-uESE-extra retained their positions. The reversal was arithmetically attributable to the S8-22 reference revision, which changed the COSMO-RS minus Chemaxon absolute-error gap by 1.58 pKa units. Because remeasurement followed large prediction–experiment disagreements, these observations characterize reference-version sensitivity within a selected subset. They do not establish benchmark-wide ranking instability or the superiority of either experimental-reference version. The case supports retaining explicit reference-version provenance alongside benchmark results. The numerical analysis uses preserved SAMPL8 inputs and does not regenerate pKa predictions. The package documents the source versions, fixed SST assignments, and stable prediction keys needed to reproduce this selected-subset comparison. Source repository: https://github.com/shomiyazaki-ai/sampl8-pka-reference-sensitivity License scope: N1 code is licensed under MIT. N1 manuscript, Supplementary Information, figures, tables, documentation, and derived data are licensed under CC BY 4.0. Upstream files retain their original MIT or CC BY 4.0 notices and attribution. See LICENSE, LICENSE_CC-BY-4.0.txt, and upstream/ATTRIBUTION.md for the file-level scope. This work received no specific external funding.