Fidelity-gated synthetic SCM that stress-tests probability-of-default modeling under selective labels and reports the model's honest operating frontier. The do() oracle real lending data can't be. Grades g-computation vs naive conditioning against planted ground truth. sklearn-only; 66 tests; fully deterministic.
calibration synthetic-data positivity credit-risk g-computation do-calculus structural-causal-model inverse-propensity-weighting selective-labels unobserved-confounding
-
Updated
Jun 13, 2026 - Python