Version selection was completed on 6 August 2026. Registry publication and digest verification were completed on 7 August 2026.
This record covers direct pipeline tools, workflow runtimes, custom-image bases,
and explicitly installed scientific libraries. Transitive packages are captured
by immutable image digests and the custom-image inventories. Database/reference
assets, including the existing Sandpiper and GTDB filenames, are unchanged and
deferred to the future db_prep work.
software.tsvis the direct-version source of truth.images.tsvmaps every configured process image to its digest, architecture, processes, direct versions, build date, and inventory.packages/containsconda list --explicitandpip freeze --alloutput exported from each published custom-image digest.
All nine custom images were cross-built and loaded locally as linux/amd64
images with Docker Buildx. Their common base is
condaforge/miniforge3:26.3.2-3@sha256:95195a50e20f7929a31a68ea407cefe117423258639efb6b66f0518e4b86f8ee,
the registry's linux/amd64 platform digest. New versioned tags were used; no
existing tag was overwritten.
The custom Myloasm image combines Myloasm 0.6.0 with mylotools 2.1.0 so
final_contig_graph.gfa can be reconciled with the polished, filtered assembly
before the pipeline publishes assembly.gfa.
COMEBin retains PyTorch 1.13.1/CUDA 11.7. VAMB and LorBin GPU retain PyTorch
2.6.0/CUDA 12.4. LorBin uses upstream commit
e35c65b3a97fe225b04dc0beaddb85fcc4a1af7c. These compatibility stacks were not
loosened while moving from Mambaforge to Miniforge.
Each custom image passed its Dockerfile build assertions and local
version/import/CLI checks before publication. After publication,
verify_custom_images.sh pulled all nine by
their tag-and-digest reference and repeated the checks. The mapping image also
passed a tiny Bowtie2/SAMtools/minimap2 read-to-reference smoke test. Published
manifest digests were independently resolved with Docker Buildx and match
images.tsv.
test_container_provenance.py
rejects mutable or malformed refs, missing digests, undocumented process images,
direct-version mismatches, missing custom inventories, and workflow-runtime
version drift.
GPU validation is limited to image build, imports, package consistency, and
PyTorch CUDA build metadata. No GPU hardware execution or numerical validation
is claimed. A full reference-dependent scientific workflow was not run; database
preparation and corresponding end-to-end validation remain outstanding for
db_prep.