Version 1.1 notes

Version 1.1 adds optional parameter-only export for clusterCov and LocalCov, with per-component disclosure diagnostics. The basic constructors, fit(), synthetic and diagnostic methods remain available. Model export does not establish disclosure clearance; see Model export and loading.

Corrections in 1.1

  • LocalCov accepts an optional trailing random_state. Explicit integer seeds now control both kNN anchor selection and NumPy/SciPy callback draws, with the same results across worker counts in the same numerical environment. Seeded parallel calls use isolated loky processes, not a threading override. The caller’s NumPy RNG is restored; custom callbacks with independent random generators are outside this guarantee. Omitting the seed retains legacy behaviour.

  • Count arguments accept NumPy integers and reject booleans and fractions.

  • Failed refits clear stale synthetic results and diagnostics instead of leaving an earlier fit marked as successful.

  • Locally constant decimal values are preserved exactly, including without clipping.

  • Fitted mixture parameters are read-only. Callers cannot silently change sampling or exported parameters independently of the fitted factors and diagnostics.

  • Covariance validation and factorization use variance-scaled units. Invalid small-scale blocks are no longer hidden by a large variance elsewhere. Constant dimensions require exactly zero covariance rows and columns.

  • Real-valued input is required; complex source values are rejected before conversion can discard their imaginary parts.

The version-1 export schemas are unchanged. The corrected factorization can change exact seeded samples relative to previous local builds. Direct and reloaded samples still agree within the same build and numerical environment; historical paper results are not promised to be bitwise reproducible. The replication materials and saved research results have not been changed or rerun.

Version 1.0 notes

Version 1.0 is a stability release. It keeps the existing public API while making the package safer for PyPI users.

Highlights

  • Public constructors and method names are unchanged.

  • Input data is checked when a resampler is created.

  • Missing numeric values are still filled with medians during fitting.

  • Non-numeric columns fail early with clear error messages.

  • Distance diagnostics handle small samples and large K values.

  • Local covariance sampling is regularized for small or singular neighborhoods.

  • Variables that are constant inside a k-neighbor subsample or cluster are preserved exactly in the generated synthetic values.

  • Unit tests cover the main resamplers, validation, metrics, and helper edge cases.