Executed validation
Validation used Python 3.12 and a CPU. reproducibility/environment.json records the installed scientific packages. These checks verify implementation and data isolation; benchmark-scale accuracy is reported only where an existing experiment record supplies it.
Measured-data workflow
The complete quickstart ran from the included Norman fixture: 920 measured cells, 20 noncontrol conditions, and 33,694 source genes. Training-only selection retained 201 variable genes and 10 principal components. The small model trained for two epochs and reloaded its validation-selected checkpoint.
The run completed forward evaluation, independent-outcome nomination evaluation, response export, exhaustive catalog ranking, three-step embedding guidance, greedy gene selection, and CSV/PDF/PNG result export. Four held-out conditions were evaluated. The target query came from measured test cells. results/validation/ contains the saved per-condition outputs and training history. checkpoints/example/best.pt and fixtures/example_cache/ preserve the matching trained example.
A separate one-epoch run exercised the population-loss branch with weight 2, two sampled conditions, and eight cells per condition. All six implemented baselines completed forward and nomination evaluation on the same small cache. Neural baselines used two epochs and 32 hidden units for this execution check.
Mathematical and data-isolation checks
The regression suite checks diagonal identity, permutation-invariant intervention pooling, the composition loss's corresponding source time, finite-difference reward gradients, agreement of NumPy and PyTorch MMD, retrieval censoring, gene holdouts, disjoint partitions, training-only PCA fitting, invariance of fitted features to held-out expression changes, and invariance of model ranking to reserved outcome changes. All 11 checks passed.
Dataset and integration scope
The Norman source download passed its published checksum. The Replogle source metadata and matrix shape were read from the published HDF5 file. Fetching response rows then stopped with HTTP 429, so Replogle training was not executed. Preparation supports its intervention labels and configurable context columns; use --batch-col batch --celltype-col cell_line.
Corrected full-scale, multi-seed runs remain pending. The GEARS and CellFlow comparisons need shared-protocol adapters, and independent interaction-residual analysis needs its evaluation implementation. These are listed in docs/experiments.md. The example checkpoint is a workflow demonstration. Historical trained weights remain available with their original source and protocol metadata.
Archive checks
The repository ZIP was extracted into a separate directory and installed into a virtual environment with pip install --no-deps --no-build-isolation. The environment used the recorded scientific dependencies already installed on the host. The imported package resolved to that environment's installed copy. All 11 tests passed there, and the complete quickstart ran again from the extracted measured-data fixture. The minimal Overleaf ZIP also compiled after extraction. The final manuscript has eight main-text pages and 17 pages including statements, references, and appendices; no overfull boxes or undefined references were reported.