Estimating Causal Effects from Tabular Data with Assumption Gates and Automated Estimator Selection for Trustworthy Artificial Intelligence

Causal analysis can guide decisions by estimating how interventions change outcomes and making the assumptions behind those estimates explicit. We present the Causal World Foundation Model, a framework for answering causal questions from tabular data. It checks declared conditions, combines classical estimators, and calibrates uncertainty. Effect queries receive an estimate with diagnostics or an explanation for refusal. A pretrained neural router serves as a baseline for comparing estimator selection and aggregation policies. The evaluation covers treatment effects, observed regimes, and network interference, alongside a semi-synthetic test using energy sensor covariates. In confirmatory experiments, a shallow selector and a convex stack improve accuracy over the evaluated deep router. Stress tests demonstrate consistent refusal when declared conditions fail, while causal interpretation remains conditional on truthful and complete declarations. A separate audit clarifies the strengths and limitations of empirical support checks. Together, these findings support a practical approach to transparent causal artificial intelligence through flexible estimator selection and clear reporting of assumptions, uncertainty, and decision reasons.