# PDR 000: Build an open, evidence-bounded pathology research programme

- HTML version: https://robbiepalmer.me/projects/pathologize/pdrs/000-open-evidence-bounded-research
- Project: Pathologize
- Status: Proposed
- Opened: 2026-11-06

## Context

Modern computational pathology now has public whole-slide datasets, foundation models, processing toolkits and
benchmark suites. An individual project is unlikely to outspend the groups training those models. It can still
contribute careful data lineage, reproducible comparisons, reliable software and evidence about where methods
fail or transfer.

Dataset curation, annotation design, identity resolution and benchmark construction are research contributions
in their own right. They often receive less attention than a new neural architecture even when better cohorts,
labels and ontologies remove a larger barrier to useful cancer research.

The work could also demonstrate relevant engineering ability to open-source maintainers and potential employers.
That incentive creates a risk of optimising the project for impressive claims rather than useful or reproducible
research.

## Proposed decision

Build Pathologize as an open, reproducible computational-pathology research programme. Treat curated datasets,
new annotation layers, ontology work, benchmark design, evaluation and software improvements as first-class
outputs where their licences permit publication. Publish enough provenance, code, parameters, resource
measurements and failed-result context for another researcher to assess each claim.

Keep the project explicitly research-only unless a later decision establishes the evidence, governance and
regulatory boundary for another use. Do not describe benchmark performance as clinical validity. Treat portfolio
value as a consequence of useful public work, not as a substitute for it.

## Consequences

The project can make a credible contribution without training the largest model. A reusable cohort, annotation
set or benchmark can be the main research result. Public artifacts and honest negative results may be valuable
even when they do not produce a new state of the art.

Open publication requires continuing licence, privacy and disclosure review. Some useful datasets, weights or
results may remain referenced but undistributed. Evidence production adds work to every experiment and may make
progress appear slower than a notebook-only reproduction.

## Evaluation

Accept the proposal after one external user can reproduce a published baseline from its public instructions and
trace its score to source data, code, weights and parameters. Reject it if the project mainly produces screenshots,
leaderboard numbers or portfolio claims that cannot be independently assessed.

## Related records

- Project: [Pathologize](https://robbiepalmer.me/projects/pathologize.md)
- Evidence: [Patho-Bench public tasks and dataset splits](https://huggingface.co/datasets/MahmoodLab/Patho-Bench)
- Evidence: [TRIDENT whole-slide processing toolkit](https://github.com/mahmoodlab/TRIDENT)

---

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