Changelog

All notable changes to industrialstats are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Nothing yet.

[0.3.0] - 2026-09-16

Added

• Genuine definitive screening designs backed by conference-matrix construction, algebraic property tests, and a published Jones-Nachtsheim reference case.
• Regular two-level factorial blocking with explicit block generators, visible confounding semantics, treatment-based block assignment, and canonical published blocked-design validation.
• Canonical two-level factorial effect estimation shared across design and analysis APIs, including higher-order interaction support.
• General factorial model-term generation with hierarchical saturated or truncated interaction structure and correct mixed-level degrees of freedom.
• Split-plot experimental-unit semantics, restricted randomization, classical whole-plot/subplot error strata, mixed-model inference, and expected-mean- square validation.
• Expanded Plackett-Burman catalogue support with NIST reference validation and foldover guarantees that separate main effects from two-factor interactions.
• Property-based validation for definitive-screening and simplex-lattice mixture designs.
• Published Cornell/NIST three-component simplex-lattice mixture reference case.
• Monte Carlo validation for factorial-effect recovery, ANOVA Type I error, response-surface coefficient recovery, split-plot variance components, and balanced ANOVA under mild heteroskedasticity.
• Independent statsmodels reference checks for two-sample t-test and one-way ANOVA power calculations.
• Structured DataExcept coverage for additional configuration, CLI, transform, CRD, and split-plot response-data boundaries.
• Milestone 2 and Milestone 3 validation audits documenting the package's current evidence base and remaining independent-reference gaps.

Changed

• Factorial blocking now derives block membership from defining treatment contrasts rather than row position.
• Split-plot analysis now treats replicated whole plots as distinct experimental units and preserves the intended two-stratum error structure.
• The package's validation policy now emphasizes hand-derived identities, published references, property tests, and repeated-sampling studies rather than shape/run-count checks alone.
• Data-loading and transformation boundaries expose more structured exception metadata while preserving native mathematical errors where appropriate.

Fixed

• Corrected definitive-screening semantics that previously did not implement a genuine DSD construction.
• Removed factorial-effect sign/scale inconsistencies between design and analysis paths.
• Prevented split-plot and blocking logic from depending on row order where treatment or experimental-unit structure should determine the result.
• Hardened configuration parsing and optional-dependency failures with explicit structured exception types and preserved causes.

[0.2.0] - 2026-09-04

Added

• PEP 561 py.typed marker, so the shipped type annotations are visible to type checkers in downstream projects.
• Documentation site built with MkDocs Material and mkdocstrings, published to GitHub Pages.
• SECURITY.md, issue forms (including a statistical-correctness report), a pull request template, and CODEOWNERS.
• Test coverage measurement with an enforced floor, plus CodeQL and pip-audit scanning in CI.
• CI now runs on pushes to main and covers Windows and macOS in addition to Linux.

Changed

• Replaced black, isort, and flake8 with Ruff for both linting and formatting. The previous flake8 configuration silenced real defect classes (unused imports and variables, bare except, star imports); these are now enforced.
• Enabled real type checking. mypy previously ran with ignore_errors = true, which disabled it entirely. Six numerically dense modules carry documented, shrinking debt via per-module overrides; everything else is clean.
• zip() calls over design matrices now pass strict=True, so a length mismatch raises instead of silently truncating.
• The CLI reports a stable program name in help output rather than inheriting it from sys.argv[0].
• Example scripts use numpy.random.Generator instead of the legacy global numpy.random.seed API.

Removed

• BREAKING: industrialstats.visualizations no longer re-exports its third-party imports. A star-import plus a computed __all__ made np, pd, plt, sns, go and stats part of the package's public API; the module now exports only ExperimentPlotter and ResponseSurfacePlotter. Import those libraries directly instead of via industrialstats.visualizations.

Security

• Raised the pytest development dependency past GHSA-6w46-j5rx-g56g (predictable /tmp/pytest-of-{user} handling on UNIX, allowing a local user to cause denial of service or possibly escalate privileges). The constraint was ^8.0.0, which pins to a range that is affected in its entirety, and did not match the 9.x that CI already installed. It is now >=9.0.3,<10, matched by minversion and by an explicit floor in the workflow install steps. pytest is a development dependency and is never installed by users of the package.

Fixed

• Eight modules had from __future__ import annotations placed above their module docstring, which left __doc__ as None and would have produced empty API documentation.
• DesignValidator.check_confounding no longer emits a divide-by-zero RuntimeWarning on perfectly confounded designs, which is the case it exists to report; it returns an infinite VIF as expected.
• Guarded design_matrix access in the CRD, RCBD, and screening designs, which previously raised AttributeError when called before generate_design().
• CompletelyRandomizedDesign validates response columns with pandas.api.types.is_numeric_dtype, which handles nullable extension dtypes that numpy.issubdtype rejects.
• Mutable default arguments in PowerAnalysis (factor_levels, powers) are no longer shared across calls.
• Exceptions raised while wrapping model-fitting failures now chain the original error with raise ... from.
• ModelDiagnostics.assumption_tests no longer depends on SciPy's deprecated implicit Anderson-Darling behaviour. SciPy 1.17 requires an explicit p-value method and removes the critical-value tables in 1.19; the test now requests an interpolated p-value where that is supported and falls back to the tables on older SciPy. The reported anderson mapping gains a p_value key alongside critical_value_5pct; whichever the installed SciPy cannot supply is None.
• ModelFitting.regularized_fitting no longer passes alphas=None explicitly to scikit-learn, which is deprecated and removed in 1.9. The argument is now omitted when unset, preserving the automatic grid on every supported version.
• DesignValidator.check_confounding also suppresses the singular-matrix conditioning warning newer statsmodels emits for perfectly confounded designs, which is the case the function exists to report.
• The performance regression benchmark compared two adjacent timings of the same call, so a single descheduled run failed it. It now warms up and takes the minimum of several repetitions on each side.

0.1.0 (2026-09-04)

Features

• add DataExcept dependency (852723c)
• add DataExcept file boundaries (55f9156)
• add structured CSV loading boundary (df894b0)
• export structured CSV loader (3762fd6)
• route CLI file I/O through DataExcept boundaries (3576008)
• wrap export boundary failures with DataExcept (ba1948d)

Bug Fixes

• preserve ANOVA CSV output format (0a185d8)
• preserve configurable CSV index semantics (4fe6169)

Documentation

• add changelog for automated releases (7ce4c6d)
• add PyPI and Zenodo release checklist (471fdb3)
• document automated release flow (82d6890)
• install development tools in setup instructions (f148493)
• let release please own changelog entries (e36a4af)
• prepare README for package release (dc3ddcc)
• reflect active DataExcept runtime integration (81c9abe)
• refresh README and correctness-first roadmap (efbadb8)
• refresh README with current DOE capabilities (7d5d1cd)
• replace stale roadmap with correctness-first plan (7bb2a91)
• simplify release flow (feb4999)