Imperfect Simulator Interventions: AI for Science

Computational simulators are increasingly used as surrogates for interventional experiments in causal discovery, but real-world simulators inevitably introduce variable omission, intervention noise, and structural misspecification. We show that a two-layer strategy---separating skeleton discovery (the undirected graph of edge presence) from edge orientation---is the key to robust causal structure recovery under such imperfect simulators. Our pipeline, QCSI, uses observational data alone for skeleton estimation (via PC or a nonlinear kernel test), and reserves simulator-generated intervention data exclusively for edge orientation (via MMD asymmetry). This separation prevents intervention noise from contaminating the skeleton, which we prove degrades as $O(k/d)$ under $k$ latent confounders (Theorem~\\ref{thm:discontinuity} and Corollary~\\ref{cor:f1_decay}). A nonlinear kernel variant (QCSI-KCIT) achieves the best performance across all four evaluation domains. Orientation accuracy 1.000 ($k \\leq 3$) and 0.960 ($k = 6$) on synthetic benchmarks, and the highest skeleton F1 on real data (biology: 0.621, engineering: 0.89--0.90). Even under severe simulator imperfection, the pipeline retains its advantage over all baselines (orientation accuracy 0.638 at structural misspecification $\\delta = 1.0$, vs.\\ 0.415 for PC alone). We validate these results on real-world data from a physics simulator ($d = 6$, $n = 330$), a biological signaling benchmark ($d = 11$, $n = 7{,}466$), and an engineering time-series dataset ($d = 3$--$4$, 195 units across five sources), confirming that the two-layer strategy consistently outperforms monolithic approaches (PC, GES, FCI, ICP, IGSP) that mix observational and interventional signals.