Roadmap¶

The library grows in deliberate layers. Each layer ships with tests that prove exactness and documentation that explains the statistics, or it does not ship.

0.1: the core objects¶

Sample. The sampled units with their design information: inclusion probabilities, strata, clusters, base weights, and a unit identifier. Built from a design in a simulation or from a data frame in practice.
Weights. A weight vector with optional replicate weights, a record of the scheme that made them, and methods that transform both together.
Estimands as functions of weights. Totals, means, ratios, and domain versions, each returning a point estimate and, when replicate weights are present, a variance and an interval.

0.2: variance and adjustment¶

Replicate-weight construction for common designs (jackknife, bootstrap, Rao-Wu), with parity tests against R's survey and svrep.
Nonresponse adjustment within weighting classes.
Calibration to known totals.

0.3: model-assisted estimation¶

Generalized-difference and model-assisted estimators that accept any learner with a scikit-learn style fit and predict.
Cross-fitting and leave-one-unit training routes.

Later¶

Two-phase designs and follow-up subsampling.
Conformal prediction for finite-population inference.
Integrations that put modern prediction models, including large language models, to work inside a design-based framework with honest uncertainty.

Non-goals¶

Wrapping R. The library is native Python with its own tested core.
A modeling framework. Learners plug in; they are not provided here.
Questionnaire or data collection tooling. This library starts where the microdata ends.