Interventional Consistency of Latent Predictors in Causal Simulations with Text-Derived Causal Statements

A predictor that only captures correlations fails when a variable is intervened upon. We test whether a joint-embedding predictive architecture (JEPA) trained on observational data predicts the correct latent outcome under interventions in simulated causal graphs with known mechanisms. We also test whether causal statements extracted from text can condition the predictor. A pilot arm confirms that confounding biases the correlational predictor; the JEPA results are pending. Preprint. Results in this version come from small-scale pilot experiments; the full experimental protocol is described in the text and is pending.