Configuration Reference¶

Python signatures are authoritative. config/settings.yaml mirrors defaults in machine-readable form and records recommended profiles, but the package does not load that YAML automatically. Pass non-default values explicitly to constructors and functions.

Core metric¶

Parameter	Default	Used by
`epsilon`	`1e-9`	`ll`, probability-mode IMV functions
`method`	`"brentq"`	`get_w` and all IMV entry points
`bounds`	`[(0.5, 0.999999999999)]`	`get_w` only
`guess`	`0.5`	Legacy `lbfgsb` backend only
`tolerance`	`1e-9`	Legacy `lbfgsb` backend only
`chance_tolerance_nats`	`0.5`	`get_w` below-chance boundary policy

High-level IMV functions expose epsilon, tolerance, and method. They do not expose custom weight bounds or below-chance tolerance; call ll and get_w directly if an audit requires those low-level controls.

`BinaryIMV`¶

Parameter	Default	Meaning
`split_method`	`"kfold"`	Shuffled K-fold parity mode
`n_splits`	`5`	Number of folds
`prop_test`	`0.2`	Holdout fraction in split modes
`model_type`	`"classification"`	Only accepted model type
`all_combinations_imv`	`None`	Optional precomputed coalition mapping
`random_seed`	`42`	Split random state
`n_jobs`	`1`	joblib workers across coalitions
`verbose`	`False`	Progress and summary output

Supported split methods are kfold, stratified_kfold, train_test_split, and stratified_train_test_split. The default preserves legacy unstratified behavior; stratified_kfold is usually preferable for new imbalanced i.i.d. analyses.

`MulticlassIMV`¶

Parameter	Default	Meaning
`n_splits`	`10`	Number of folds
`optional_explanatory_variables`	`None`	Use all columns except the outcome
`random_state`	`None`	Fold random state
`stratified`	`False`	Preserve original shuffled `KFold`
`verbose`	`False`	Print fold summaries

Set stratified=True and a fixed random_state for a new ordinary multiclass analysis unless a prespecified design requires otherwise.

`AblationIMV`¶

The constructor defaults to random_seed=42 and selects devices in CUDA, MPS, CPU order. Training defaults are three epochs, learning rate 2e-5, torch.optim.Adam, no scheduler, constructor seed, and verbose output.

Prediction frames default to columns True Label and Positive Probability. Pass target_column and prob_column when using another schema.

Plotting¶

Shared heatmaps default to size (6, 6), the custom imv navy-to-red colormap, and three-decimal cell annotations. The style uses Helvetica where available, then Nimbus Sans and DejaVu Sans as portable fallbacks. Interactive figures use 110 DPI; save_figure always emits PNG, PDF, and SVG at 800 DPI with bbox_inches="tight" and 0.04-inch padding.

Reproducibility profiles¶

config/settings.yaml records paper_parity and recommended_production profiles. They are documentation, not runtime presets. The parity profile keeps historical split choices where possible; the production profile recommends stratification, ten repeated seeds for ablation, directional matrices, and aligned test rows. Copy only the settings justified by the current analysis.