Skip to content

Provenance fields

<sample>/summary/summary.json contains the completion and scientific identity:

  • slide ID, source path identity/fingerprint, dimensions;
  • source slide_mpp and target mpp/target_mpp;
  • scientific processing signature and schema;
  • paper-figure layout version, when the current rendering contract is recorded;
  • HistoPLUS weight identity;
  • device;
  • detected-cell/class totals and explicit zero_detections;
  • deterministic tile/patch sampling summary;
  • pyramidal conversion identity; and
  • relative output paths.

<sample>/summary/run_metadata.json records settings such as:

  • source-MPP origin (override or embedded);
  • tile size/overlap/background fraction;
  • percentage and random seed;
  • model repository, immutable revision, magnification, and weight filename;
  • selected device/batches/workers;
  • overlay palette/style;
  • paper-figure layout version and portable relative figure/legend paths;
  • conversion settings; and
  • whether a token was supplied as a Boolean, never its contents.

The paper-figure layout version is deliberately excluded from the scientific worker processing signature. This separation means that direct reuse of the same persistent worker output preserves the earlier completion contract for a legacy summary without a layout version. For outputs whose completion summary records the current layout version, external paper_figures/celltypes_paper_figure_legend.txt is part of the validated contract; missing or empty text prevents completion/resume reuse.

Worker-signature reuse is not the same as top-level workflow-cache reuse. PROCESS_SLIDE stages the worker code, and Nextflow also keys that code and its configuration. A software upgrade may therefore invalidate the task and re-enter HistoPLUS inference even though the scientific worker signature is unchanged. Before a top-level resume, inspect the --resume --dry-run plan for Nextflow -resume and the original work directory. Preserve the exact software revision and work cache, or select --cpu when reuse is uncertain. To obtain redesigned artwork for a legacy result, use a new output directory or choose a deliberate rerun.

The legend records only public sample identity and portable scientific provenance; it must not contain private cohort identifiers or absolute source paths.

Workflow metadata adds:

  • Git/software version or commit when available;
  • immutable container digest;
  • Nextflow report, trace, and timeline;
  • discovery manifest/fingerprints; and
  • authoritative aggregation roster/audit.

The public one-slide quickstart additionally records Zenodo record 21466410, dataset DOI 10.5281/zenodo.21466410, dataset-matched release v0.4.0, MDS/TIFF checksums, source MPP 0.261780, conversion levels L0/L2, command provenance, and the sampling seed.

Secrets and model weights are external resources. Token contents and weight files are never copied into result provenance.

Weight-identity privacy migration

Current outputs retain the authorized weight filename, byte size, and SHA-256 needed for identity checks. They intentionally omit its absolute path, device/inode numbers, and filesystem timestamps. Older outputs may contain those location-specific fields; downstream readers must accept both forms and must not require or republish the removed fields.