Improving RCT-Based Treatment Effect Estimation Under Covariate Mismatch via Calibrated Alignment — ThinkLLM