trustworthy_survey: a benchmark and calibration toolkit for machine-learning predictions under complex survey sampling

It provides (i) an estimand-first benchmark protocol with a four-rung ground-truth validation ladder (synthetic, semi-synthetic, official-anchor, and cross-cycle/cross-survey transport); (ii) a post-hoc calibration layer enforcing exact agreement with arbitrary weighted margins, with exactness theorems and a first-order error-propagation law;(iii)design-consistent uncertainty quantification (Fay-BRR with generalized unbalanced-strata replication, averaged-SE intervals); and(iv)a documented kill-list of disproven approaches.

All results are generated from public NHANES and NHIS microdata; the archive includes the analysis scripts and all numerical outputs (CSV) behind the associated preprint (ChinaXiv T202609.00070, under review).