# PDR 001: Use one versioned pathology data registry

- HTML version: https://robbiepalmer.me/projects/pathologize/pdrs/001-one-versioned-pathology-data-registry
- Project: Pathologize
- Status: Proposed
- Opened: 2026-11-06

## Context

Pathology experiments reuse the same large source slides, case identities, scanner metadata and public labels.
If each model pipeline downloads and reshapes its own copy, licence records and patient boundaries drift. Derived
tiles can also become mistaken for independent samples or canonical data.

Quality control, tissue segmentation, cell analysis, tiling, slide encoding and biomarker prediction need
different transformations and leakage-aware partitions. Putting every derived artifact into one pipeline would
couple unrelated experiments and make the source dataset hard to update.

## Proposed decision

Maintain one authoritative, versioned pathology data registry for source artifacts and source-native metadata.
It will own provenance, rights, checksums and stable patient, case, specimen, slide and image identities.

Independent downstream projects will consume a pinned registry revision and own their task-specific eligibility,
partitions, transformations, tiles, embeddings, models and scorecards. Raw slides remain canonical source
artifacts. Derived tiles and features remain reproducible outputs of the project that defines their parameters.

Preserve source-native labels. A downstream project may add a more detailed annotation layer or migrate the same
material to an incommensurable ontology. Version the new ontology, annotation protocol, annotator provenance,
agreement evidence and any partial cross-dataset mapping as derived artifacts. Record categories that cannot be
mapped instead of forcing them into a universal label scheme.

## Consequences

The registry gives every experiment a shared provenance and leakage boundary while allowing pipelines to evolve
independently. Storage can deduplicate immutable artifacts without giving several projects ownership of them.

Consumers must manage pinned imports and deliberate updates. Identity resolution and licence review become
registry responsibilities. Cross-project changes require compatibility checks, and the graph must remain acyclic
enough that a registry update can be understood without reproducing every downstream experiment.

## Evaluation

Accept the proposal when two downstream projects reproduce from the same pinned registry revision, derive
different artifacts without modifying the source data, and report the same patient and slide identities. A
registry update must show which consumers are stale without silently changing their results.

## Related records

- Project: [Pathologize](https://robbiepalmer.me/projects/pathologize.md)
- Evidence: [DVC data registry and import contract](https://dvc.org/doc/command-reference/import)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt