CLI
Install in editable mode before using the command-line interface:
python -m pip install -e ".[dev]"
Simulate
icg-cast simulate --n 1200 --months 72 --seed 7 --outdir outputs/demo
Writes a synthetic cohort, metadata, and optional trajectory plots.
To sample one coefficient-prior realization for the whole cohort:
icg-cast simulate \
--n 1200 \
--months 72 \
--seed 7 \
--coefficient-mode prior_sample \
--coefficient-seed 42 \
--outdir outputs/demo_uncertainty
Make Demo
icg-cast make-demo --n 120 --months 72 --seed 7 --outdir outputs/demo
Runs the reproducible demo workflow:
simulatetrainevaluategraphplot generation
It writes demo_manifest.json alongside the generated cohort, metrics, model
bundle, graph exports, and PNG plots.
make-demo accepts the same --coefficient-mode and --coefficient-seed
flags as simulate.
Train
icg-cast train \
--cohort outputs/demo/synthetic_icg_cohort.csv \
--outdir outputs/demo \
--seed 7
Writes:
model_metrics.csvpermutation_importance.csvcounterfactual_tests.csvbiological_coherence.csvmodel_bundle.joblibmodel_card.mdmodality_auc.pngunless plots are disabled
The saved bundle contains the best multiomics_plus_qAOP baseline model and
the held-out split indices used during training.
Evaluate
icg-cast evaluate \
--cohort outputs/demo/synthetic_icg_cohort.csv \
--model outputs/demo/model_bundle.joblib \
--outdir outputs/demo
Writes evaluation metrics, calibration diagnostics, counterfactual tests, biological-coherence summary, and a model card. By default the command evaluates on the held-out split stored in the bundle.
Graph
icg-cast graph --outdir outputs/demo
Writes:
icg_theory_graph.graphmlicg_theory_graph_edges.json
Bench
icg-cast bench list
icg-cast bench info linear_lowhet
icg-cast bench run --cohort linear_lowhet --variant sign_constrained_augmented --seed 7
Benchmark commands run synthetic data-generating-process variants with known mechanistic structure for recovery and intervention-conformity experiments.