The three datasets are public under their own licences and are not
redistributed here. Place them in this directory so that the frozen programs
find them at the paths they were run with; data/source_registry.json holds
the authoritative URLs, sizes and digests, and scripts/verify_hash_registry.py
re-hashes whatever is present.
work/data/
osf_938rs/ DAFD deposit (Lashkaripour et al. 2024), https://osf.io/938rs/
Comprehensive_normalized.xlsx
Generalizability_data_normalized.xlsx
Final_Comprehensive_plus_Generalizability_data_normalized.xlsx
Raw_SE_dataset.xlsx
zenodo_17867478/ Smart Microfluidics liposome dataset (Lavagna et al.; CC BY 4.0)
microfluidics_dataset.zip
README.pdf
microfluidics_dataset/ the archive extracted in place (raw_data/*.csv are the bound files)
The two Chagot MATLAB tables (Zenodo 7055018, CC BY 4.0) go to
work/microfluidic_reliability/blind_raw/ instead, because the V1 custodian
lock in native_physics_v2.py verifies them at that frozen path:
work/microfluidic_reliability/blind_raw/
matTrainingDataSet.mat
matTestDataSet.mat
Download and verify:
python3 scripts/fetch_public_data.py --dataset chagot_zenodo_7055018 --destination work/microfluidic_reliability
python3 scripts/fetch_public_data.py --dataset smart_microfluidics_zenodo_17867478
unzip -q work/data/zenodo_17867478/microfluidics_dataset.zip -d work/data/zenodo_17867478/
# OSF objects have no stable direct download URL recorded here: download the four
# workbooks from the project page into work/data/osf_938rs/ and verify them:
python3 scripts/fetch_public_data.py --dataset dafd_osf_938rs --verify-onlyExpected SHA-256 digests are in data/source_registry.json and in the
manuscript's Appendix D (hashes/appendix_d_registry.json).
The V2, W3–W5 and post-hoc programs read the Chagot rows through the V1
custodian lock (blind_artifacts/development_labeled.csv,
blind_features.csv, sealed_blind_labels.csv), and the post-hoc programs
read the D3 tables (d3_blind_artifacts/development_labeled.csv,
sealed_*_features.csv). These re-cuts of public rows are not committed; the
custodian programs regenerate them byte-for-byte (the digests are verified
against the committed custodian manifests, and both manifests are themselves
reproduced byte-identically):
python work/microfluidic_reliability/src/blind_custodian.py \
work/microfluidic_reliability/blind_raw/matTrainingDataSet.mat work/microfluidic_reliability/blind_raw/matTestDataSet.mat \
--source-url 'https://zenodo.org/records/7055018/files/matTrainingDataSet.mat?download=1' \
--source-url 'https://zenodo.org/records/7055018/files/matTestDataSet.mat?download=1' \
--source-md5 c5dd5f3fbfd7c218cea734b205ae502e --source-md5 a6c41bba18e4f232e7f25e2207fd183c
cd scaleup && python -m w3.d3_custodian --download-dir ../work/data/zenodo_17867478 \
--binding D3_BINDING.json --protocol D3_BLIND_PROTOCOL.md --output-dir ../work/microfluidic_reliability/d3_blind_artifactsBoth commands write into the committed artifact directories; git ignores the
regenerated tables and the sealed label files (mode 0600). Compare the written
custodian_manifest.json with the committed one (git diff must be empty) and
run python3 scripts/verify_hash_registry.py, which then reports the
regenerated tables as verified.
work/data/
h_A1/PMC11834948_supplementaryFiles.zip Hormozinezhad et al. 2025 ESI (Europe PMC bundle; CC BY 3.0)
h_R2A1/Data for FrameworkA.xlsx Wang et al. 2024, Zenodo 10526441 (CC BY 4.0)
h_R2A2/LNP_Atlas_DB_202509_v1.csv LNP Atlas, Zenodo 17243733 (CC BY 4.0)
upstream/dafd-website/ git clone of CIDARLAB/dafd-website at commit 4902f8ce (fidelity protocol only)
python3 scripts/fetch_public_data.py --dataset hormozinezhad_2025_esi_A1
python3 scripts/fetch_public_data.py --dataset wang_2024_zenodo_10526441_R2A1
python3 scripts/fetch_public_data.py --dataset lnp_atlas_zenodo_17243733_R2A2
git clone https://github.com/CIDARLAB/dafd-website work/data/upstream/dafd-website && git -C work/data/upstream/dafd-website checkout 4902f8cescaleup/w6/h_download.py is the program that produced the header-only intake
manifests under scaleup/h_intake/; it downloads the same files and prints no
cell value. The R10 custodians then write work/microfluidic_reliability/h_<code>_artifacts/
(development table, sealed features, sealed labels with mode 0600) whose
digests are in the released custodian manifests:
cd scaleup && python -m w6.h_custodian --binding validation_plan/H_BINDING_A1_v2.json --protocol validation_plan/H_PROTOCOL_v1.md \
--zip ../work/data/h_A1/PMC11834948_supplementaryFiles.zip --output ../work/microfluidic_reliability/h_A1_artifacts
cd scaleup && python -m w6.h_custodian2 --binding validation_plan/H_BINDING_R2A1.json --protocol validation_plan/H_PROTOCOL_v2.md \
--source "../work/data/h_R2A1/Data for FrameworkA.xlsx" --output ../work/microfluidic_reliability/h_R2A1_artifacts
cd scaleup && python -m w6.h_custodian2 --binding validation_plan/H_BINDING_R2A2.json --protocol validation_plan/H_PROTOCOL_v2.md \
--source ../work/data/h_R2A2/LNP_Atlas_DB_202509_v1.csv --output ../work/microfluidic_reliability/h_R2A2_artifactsThe closed-loop runners (scaleup/w7/) read the Stage 1 and Stage 2 tables
through these custodian outputs and the --data directory (work/data).