Abstract
Abstract
Somatic variant benchmarks can be difficult to interpret when analytical choices change after external truth is examined. We froze caller configurations, normalized exact allele matching, benchmark regions, consensus rules and spatial-bootstrap parameters after HCC1395/SEQC2 development and before protected HG008 evaluation. In the primary HG008 region, DeepSomatic and majority consensus achieved F1 scores of 0.8973 and 0.9136; their paired difference was concentrated in INDELs (delta F1, − 0.0631) rather than SNVs (− 0.0002). Analyses prespecified before HG008 access but executed after primary evaluation linked ensemble behavior to caller-support patterns and showed variation by INDEL length and benchmark context. The study does not establish clinical validity or general transportability. It shows how prospective freezing, support decomposition and spatial uncertainty can make benchmark-specific performance more interpretable.