Prediction-Powered Inference

Valid statistical inference when most labels are model-imputed (Jordan school): the PPI rectifier (small gold-standard set corrects an ML-imputed estimate; CIs cover the population estimand regardless of model quality), PPI++ power-tuning (asymptotically never worse than classical), cross-prediction when no pretrained model exists, and active label allocation under a budget. Use whenever a CI, p-value, or regression is about to be computed partly from model predictions — "we only have confirmed labels for a fraction", "can we use the LLM's labels", "survey plus model imputation", "attrition confirmed months later", "annotate a sample and extrapolate" — and as the audit th…