validation

validation/

Reproducibility scripts + measurement artefacts behind the rustscenic README numbers. Every claim made in the top-level README and CHANGELOG has a corresponding script here.

Install validation dependencies

The rustscenic core package intentionally keeps clustering and scverse dependencies out of the default install. Most collaborator-facing validation scripts do use scanpy clustering and sklearn metrics.

From a source checkout, install the validation extra before running them:

pip install -e ".[validation]"

From PyPI (v0.4.0+):

pip install "rustscenic[validation]"

Layout

• ziegler_headtohead_2026-04-19.md - flagship real-atlas head-to-head with pyscenic on 31,602 nasopharyngeal cells (Ziegler et al. 2021).
• VALIDATION_SUMMARY.md - one-page index of every measured number.
• fresh_install_proof.md - the 2026 "what happens in a cold venv" test.
• figures/ - the three headline figures (tool-validation only; biology figures live in the companion case-study).
• reference/ - pinned python:3.11-slim Docker + run_reference.py for regenerating the cached arboreto/pyscenic outputs we compare against.
• baselines/ - small committed parquets (top-10k edges) used as CI-smoke-test inputs to compare.py.
• validate_*.py, *_pipeline.py, etc. - the driver scripts.
• compare.py - the CLI tool CI calls.
• agentic/ - fresh-container scenario runner for structured install, quickstart, guardrail, and external-tester-style validation reports.

⚠️ Paths are historically hardcoded

Most of the validate_*.py scripts hardcode absolute paths to the author's filesystem (/Users/ekin/projects/bio/rustscenic/validation/reference/data/...). These are the original research scripts committed for reproducibility, not user-facing entry points - if you want to regenerate numbers on your own environment, you will need to adapt the paths. The rustscenic package itself (GitHub Release wheel / source install today, PyPI once live) does not contain any hardcoded paths.

For user-facing pipelines, see examples/pbmc3k_end_to_end.py in the repo root - that script downloads its own data and has zero hardcoded paths.

Regenerating the Docker reference

cd validation/reference
docker build -t rustscenic-reference .
docker run --rm -v $(pwd)/data:/data rustscenic-reference

This produces the pyscenic/arboreto cached outputs that the rest of the validation scripts compare against. Pinned to pyscenic 0.12.1 + arboreto 0.1.6 + dask 2024.1.1 + distributed 2024.1.1 + pandas 2.1.4 - the last versions where pyscenic's own stack cohabits cleanly in a single env.