Found two separate (pre-existing, unrelated to my SAVANA PR) issues in lrsomatic while testing with real patient data, flagging for whoever owns PREPARE_REFERENCE_FILES/model selection: 1. No HAC entries in the Clair3/ClairS model maps clair3_modelMap / clairs_modelMap in workflows/lrsomatic.nf only have entries for ..._sup@v4.x / v5.x basecaller models. Real data basecalled in HAC mode, e.g. dna_r10.4.1_e8.2_400bps_hac@v4.3.0, which is what our actual data uses, has no match, so auto-detection silently resolves to nothing. 2. Manual clair3_model override breaks model-download staging when auto-detection doesn't resolve In subworkflows/local/prepare_reference_files.nf52 def id_new = basecall_model_meta ? clair3_modelMap.get(basecall_model_meta) : basecall_model_meta def model = (!meta.clair3_model || ...) ? clair3_modelMap.get(basecall_model_meta) : meta.clair3_model model (used for the actual download URL) correctly falls back to the samplesheet's explicit clair3_model override. id_new (used as meta.id for UNTAR's output directory name) does not, it's only ever derived from the map lookup. So when a sample's basecall model isn't in the map (issue #1, or just missing metadata) and you supply an explicit override, the download itself succeeds but UNTAR crashes: mkdir: missing operand because the staging directory name comes out empty.