Changelog

All notable changes to bossanova are documented here.

0.1.0.dev25 (2026-03-09)¶

• Bug Fix use : separator in rho keys for correct effect name parsing

• Bug Fix correct plot_ranef() guard and attribute for varying effects

• Bug Fix update grouseticks_three_term_correlated parity to validate structure only

• Bug Fix add explicit per-factor RE dims for multi-factor mixed models

• Bug Fix prediction bootstrap now replays full prediction path

• Bug Fix remove unsupported BCa bootstrap for mixed models

• Bug Fix use exact finite-sample permutation p-value formula

• Bug Fix update container diagram script for split state modules

• Bug Fix include percentile CIs in permutation params DataFrame

• Docs regenerate changelog

• Docs rewrite quickstart explore/predict sections with Titanic dataset

• Docs consolidate CLAUDE.md, style guide, and design constraints (80% reduction)

• Docs reorganize project root — move MyST ref to docs, remove style guide pointer

• Docs add llms.txt generation for AI-friendly documentation

• Docs add cross-type comparison grid table to compare() docstring

• Docs normalize container lifecycle docstrings

• Docs add ast-grep to code search tools, update Linear CLI gotchas

• Feature add donate_argnums support to ops.jit() for memory efficiency

• Feature add formula syntax to model.predict() for grid-based predictions

• Feature support cross-type model comparison (lm↔lmer, glm↔glmer)

• Feature enable CV comparison for mixed models (lmer/glmer)

• Feature add holdout_group for group-aware CV splitting

• Feature vmap-accelerated GLM bootstrap inference

• Feature match R/lmerTest summary display for mixed models

• Improvement deduplicate model/ tests

• Improvement replace inline inference column assertions with helper

• Improvement relocate misplaced tests to correct layer directories

• Improvement split oversized test file and deduplicate fixtures

• Improvement 3.3x faster API doc generation via caching loader (67s → 20s)

• Improvement add donate_argnums to GLM solver JIT cache

• Improvement remove unused viz modules (cognition, dag, lattice, layout)

• Improvement tighten targeted container validator contracts

• Improvement split state containers by lifecycle

• Improvement thin model/core.py back to a proper facade

• Release v0.1.0.dev25

0.1.0.dev24 (2026-03-04)¶

• Bug Fix include interaction columns in simulation DGP

• Docs regenerate changelog

• Feature add formula syntax to plot_predict and remove plot_fit

• Release v0.1.0.dev24

0.1.0.dev23 (2026-03-04)¶

• Bug Fix include pinned-condition columns in crossed EMM output grid

• Docs regenerate changelog

• Release v0.1.0.dev23

0.1.0.dev22 (2026-03-04)¶

• Bug Fix make jointtest coding-invariant with Type III L matrix averaging

• Docs regenerate changelog

• Improvement deduplicate 4 coefficient operator closures into 2 factories

• Improvement vmap-accelerated LM bootstrap and matmul permutation inference

• Improvement vectorize per-df inference loops in marginal inference

• Improvement vectorize EMM g-computation loops with batched einsum

• Release v0.1.0.dev22

0.1.0.dev21 (2026-03-04)¶

• Bug Fix use exact g-computation for response-scale GLM slopes and bootstrap

• Bug Fix response-scale GLM dispatch for permutation and crossed EMMs

• Bug Fix continuous at-values now cross with RHS conditions in explore()

• Bug Fix explore formula edge cases — 5 bugs, 6 error message gaps

• Bug Fix BCa z₀ float32 precision loss and flaky hypothesis test

• Bug Fix apply delta-method response-scale transform in at-values EMM paths

• Bug Fix reject string responses with family-specific errors for non-binomial GLMs

• Bug Fix preserve model state and thread palette through plot_explore()

• Bug Fix auto-detect condition columns in plot_explore() dispatch functions

• Bug Fix remove verified dead code and configure deadcode tool

• Bug Fix close Col constant gaps and remove orphan schema constants

• Bug Fix store boot design matrices in uncrossed AME slopes for exact bootstrap

• Bug Fix replace bare string column names with Col constants across 22 files

• Bug Fix remove silently ignored what param and wire solver through .fit()

• Bug Fix resolve layer violations and module boundary issues

• Bug Fix widen MathDisplay.to_markdown path type to accept Path

• Bug Fix reject dict-form corr in varying() instead of silently ignoring

• Docs regenerate changelog

• Docs consolidate cheatsheet model comparison around compare()

• Docs update cheatsheet

• Docs enhance tests-as-linear-models with GLM formulas and null distributions

• Docs add synced R equivalent tabs to cheatsheet

• Docs fix explore() docstring grammar and cheatsheet formula syntax

• Docs add executable distribution showcase and normalize docstrings

• Docs fix docstring inaccuracies across model, compare, and distributions

• Docs audit and clean up project knowledge docs

• Feature auto-compute profile CIs for variance components on mixed boot/perm

• Feature add pytest-cov coverage reporting with docs embedding

• Feature add semantic internal API docs (functions + containers reference)

• Improvement vectorize t-distribution calls in Satterthwaite inference

• Improvement vectorize QR vcov unpermutation with np.ix_

• Improvement split marginal/compute.py into resolve.py + co-locate contrast dispatchers

• Improvement unify distributions — facade over internal analytical system

• Improvement cache last PIRLS result to avoid redundant post-Stage 1 re-evaluation

• Improvement pre-compute ZL_dense and replace sp.diags with broadcasting in PIRLS

• Improvement cache symbolic Cholesky in Satterthwaite Richardson extrapolation

• Improvement factor out sigma from Satterthwaite Richardson extrapolation

• Improvement extract 3 implementation blocks from model class to internal/

• Improvement make FormulaStructure required in build_model_spec_from_formula

• Improvement clean up naming, dead code, and missing annotations

• Improvement reduce cognitive complexity in detect_categoricals and compute_cv_metrics

• Improvement remove dead code and fix container re-exports

• Improvement clarify sandwich no-op branches and use abstractmethod on StatefulTransform

• Improvement decompose _dispatch_marginal_core into focused helpers (complexity 68 → <15)

• Release v0.1.0.dev21

0.1.0.dev20 (2026-03-03)¶

• Bug Fix update benchmarking suite for current API

• Bug Fix drop missing-data rows before finite-difference slope computation

• Docs regenerate changelog for v0.1.0.dev19 release

• Docs add indented sub-bullets, pre-release sections, and postprocessors to changelog

• Docs consolidate public api docs and rename

• Docs update quickstart contrast coding to inline formula syntax

• Docs sync project docs with current codebase structure

• Docs add symbex recipes to Code Exploration section

• Docs regenerate GLM, LMER, GLMER benchmark reports

• Docs add cheatsheet

• Feature add post-publish steps to release script (commit, tag, changelog, docs)

• Feature standardize git history and expand changelog with commit body details

• Feature add standalone R inference benchmarking and use lmerTest for LMER

• Improvement consolidate benchmarking docs into single unified page

• Improvement default GLMER to analytical Schur vcov, eliminating post-fit Hessian cost

• Release v0.1.0.dev20

0.1.0.dev19 (2026-03-02)¶

• Bug Fix rename plot_explore units param to effect_scale for consistency

• Bug Fix relax scipy assertion tolerances and merge paired/repeated-measures docs

• Bug Fix prevent ShapeError when calling .infer(how=“cv”) after predict/plot_predict

• Docs clean up API docs rendering — details blocks, hyperlinks, TOC, and docstrings

• Docs restore function hyperlinks in distributions summary table

• Docs clean up container-overview page — lowercase title, fix mermaid, replace classDiagrams with tables

• Docs make explore/infer kwargs explicit and fix docstring rendering

• Docs model attribute descriptions and stray }} in API docs

• Docs escape pipes in code spans within API doc table rows

• Docs link explore() docstring to quickstart cheatsheet

• Docs cap API doc headings at h5 to satisfy MyST depth limit

• Docs rename docs and re-add linear models docs

• Docs add formula and explore reference tables to model docstrings

• Docs replace set_contrasts with inline formula syntax in model docstring

• Docs regenerate changelog after fix→docs commit message corrections

• Feature add hyperlinks to internal API overview tables

• Feature add to_markdown/to_csv export methods for Quarto embedding

• Feature make summary() context-adaptive for explore analyses and inference modes

• Feature add auto-generated changelog with seaborn-style pill badges

• Feature add release: commit type, backfill legacy releases in changelog

• Release v0.1.0.dev19

0.1.0.dev18 (2026-03-02)¶

• Bug Fix null-safe jaq filter in Linear CLI example

• Bug Fix remove case bootstrap stubs for mixed models

• Bug Fix remove lambda template NotImplementedError stub

• Bug Fix remove stale operations.profile from docs build

• Bug Fix handle h7+ headings in API docs post-processing

• Bug Fix propagate CVState from params CV inference to diagnostics

• Docs refactor docs and tutorials

• Docs add formula syntax, transforms, and contrast coding reference tables

• Docs complete quickstart — fix numbering, fill placeholders, resolve TODOs

• Docs update code-review.md — resolve EJO-209, EJO-212, EJO-213

• Docs move style guide to docs site, add to engineering section

• Docs relax file size limits to cohesion-based policy

• Docs overhaul internal API docs with call-chains and domain module structure

• Docs promote Internal API to top-level nav, add auto-generated overview and post-processing

• Docs fix duplicate titles, hide author block on API pages, add viz to overview

• Docs auto-discover modules in API docs build scripts

• Docs flatten API sidebar — top-level public pages, no domain grouping

• Docs fix 4,000+ MyST build warnings in API docs post-processing

• Docs consolidate build system, rename help/ to api/, auto-discover TOC

• Feature add type-based dispatch to plot_explore for means, slopes, and contrasts

• Feature normalize plot function signatures (height/aspect, prettify, col_wrap)

• Feature improve model API docs — narrow exports, expand plot docstrings, post-process headings

• Feature grayscale design matrix styling

• Feature wire in schemas, add validators, implement AIC/BIC compare

• Feature consolidate CI bands, implement RE operators, improve Welch DF errors

• Feature formula-syntax contrast coding — replace factor()/T()/S() with treatment(), sum(), helmert(), sequential(), poly()

• Feature add release orchestrator and align test/parity commands

• Feature add rerun option for stale/missing/failed suites in release Step 1

• Improvement extract infer dispatch, power analysis from model class

• Improvement formalize verb-domain naming convention, rename 20 internal functions

• Improvement consolidate resample module, remove 2,301 dead lines

• Improvement replace **kwargs with explicit params in fit, infer, marginal dispatch

• Improvement remove dead code, kill last **kwargs in internal/

• Improvement dissolve operations/, promote domain modules to internal/

• Improvement delete operations/ facade, update all consumers to direct imports

• Improvement absorb loose files into domain modules, dissolve common/

• Improvement absorb last loose files into domain modules

• Release v0.1.0.dev18

0.1.0.dev17 (2026-02-25)¶

• Bug Fix use full-precision values in MAR hypothesis tests to avoid sigfig rounding artifacts

• Feature add ModelResult proxy for consistent notebook rendering

• Feature improve plot labels, add bootstrap p-values, rename plot_mee to plot_explore

• Improvement rename weights→how, units→effect_scale, fix AME slope SE

0.1.0.dev16 (2026-02-25)¶

• Docs rewrite marginal effects tutorial with full syntax reference

• Feature add g-computation (weights=“observed”) averaging for GLM explore()

• Feature add marginaleffects R parity tests for g-computation validation

• Release v0.1.0.dev16

0.1.0.dev15 (2026-02-25)¶

• Docs work-in-progress tutorial

• Feature round DataFrame output to 4 significant figures by default

• Feature auto-display results in REPL/notebook via context-aware repr

• Feature add contrast registry, treatment/sum/helmert functions, remove revpairwise

• Feature add bracket contrast expression parser for explore()

• Feature wire bracket contrast computation into explore() pipeline

• Feature add RHS bracket contrasts for interaction and slope contrasts

• Feature add compound variable bracket contrasts for interaction cells

• Improvement replace regex explore parser with scanner+parser architecture

• Release v0.1.0.dev15

0.1.0.dev14 (2026-02-24)¶

• Bug Fix pass precision=None in compare tests that check exact numerical identities

• Docs add R/emmeans syntax comparisons to marginal effects tutorial

• Docs rewrite marginal effects tutorial with single-dataset narrative

• Feature fix binomial string response, bracket contrast syntax, consolidate formula parser

• Feature clean up explore() formula output and syntax ergonomics

• Feature add simulate(power=...) power analysis API

• Feature unify @ syntax in explore formulas, remove bare bracket syntax

• Feature accept file paths in model(data=) and add power analysis tutorial

• Feature summary() displays without print(), bump 0.1.0.dev14

0.1.0.dev13 (2026-02-23)¶

• Docs bump version to v0.1.0.dev12 in myst.yml

• Feature add precision kwarg to compare() for controlling float display

• Release publish 0.1.0.dev13 to PyPI

0.1.0.dev12 (2026-02-23)¶

• Bug Fix resample_bundle valid_mask crash with NA rows

• Bug Fix remove unused pytest imports in edge case test files

• Bug Fix align diagnostic array lengths in extract_residuals for NA data

• Bug Fix method-aware .params/.effects schemas with correct column sets and df values

• Bug Fix consolidate jax_enable_x64 to single authoritative location in init.py

• Bug Fix .resamples returns raw resampled values instead of aggregated summary

• Bug Fix CV ablation produces non-zero PRE for categorical and interaction terms

• Bug Fix bracket syntax on categorical variables in explore(); add marginal effects stress test tutorial

• Docs new penguins-based quickstart, move learn/ guides to wip/

• Docs update CLAUDE.md with fast/slow test tier documentation

• Docs miscellaneous documentation updates

• Docs new wip tutorial for inference approaches

• Docs remove legacy docs, relocate benchmarks to engineering section

• Feature Welch ANOVA, BCa bootstrap, inference fixes

  • feat: support Welch ANOVA in jointtest() and per-contrast Satterthwaite df

  • feat: route mixed model bootstrap to parametric path with BCa acceleration

  • fix: use L_matrix for permutation SE and null distribution on multi-row slopes

  • fix: group-aware CV splitting for mixed models with small group counts

  • test: cross-grid smoke test for all model types x methods x args

• Feature cluster-robust SEs for mixed models, HC→CR transparent dispatch

• Improvement absorb redteam tests into consolidated model edge-case files

• Improvement clean-up public model class and separate out guards

• Release publish 0.1.0.dev12 to PyPI

0.1.0.dev11 (2026-02-11)¶

• Release container validators, Linear migration

  • fix: add validators to 36 bare container fields

  • fix: replace Monte Carlo MVT with scipy CDF + root-finding

  • refactor: make errors and what explicit params on model.infer()

  • chore: migrate issue tracking to Linear

0.1.0.dev10 (2026-02-11)¶

• Release ECS refactor — unified model API

  • feat: unified model class replacing per-type classes (lm/glm/lmer/glmer)

  • feat: frozen attrs containers (ModelSpec, DataBundle, FitState, InferenceState)

  • feat: operations as pure functions on containers

  • feat: explore() for marginal effects with crossed slopes

  • feat: Tukey HSD/MVT multiplicity adjustment

  • refactor: extract Schur complement vcov to shared maths module

  • refactor: consolidate estimated-dispersion family set

  • docs: ECS architecture proposal and container-first workflow

  • test: comprehensive container and operation tests

0.1.0.dev5 (2026-01-31)¶

• Release MyST docs, weighted fitting

  • feat: adaptive Gauss-Hermite quadrature (nAGQ > 1) for GLMER

  • feat: BCa confidence intervals for mixed models (cluster jackknife)

  • feat: profile likelihood CIs and result_varying property

  • feat: prettify parameter and standardized plot styling

  • refactor: standardize CI band rendering across viz

  • refactor: deduplicate theta bounds and inverse link dispatch

  • docs: comprehensive marginal effects guide

  • test: consolidate recovery tests into single parametrized file

0.1.0.dev4 (2026-01-31)¶

• Release NumPy backend, Pyodide, API stabilization

  • feat: NumPy backend for Pyodide/browser compatibility

  • feat: interactive Pyodide demo page (try in browser)

  • feat: rename .anova() to .jointtest()

  • feat: weights support with column name API

  • feat: centralize DataFrame column names via Col namespace

  • refactor: move weights from init to .fit()

  • refactor: extract implementation blocks from model class

  • docs: replace legacy tutorials with marimo-converted content

  • test: Pyodide smoke tests for lm/lmer public APIs

0.1.0.dev0 (2026-01-27)¶

• Release first PyPI publish — formula rewrite, API cleanup

  • feat: vendored formula parser with Wilkinson notation

  • feat: expressions module for Polars statistical transforms

  • feat: distribution algebra with expressive composition

  • feat: compare() function for multi-model comparison

  • feat: unified lmplot-style visualization

  • feat: hypothesis property-based tests

  • refactor: split model.py and design.py into focused mixins

  • docs: MyST documentation site with API generation

  • chore: first PyPI publish as bossanova 0.1.0.dev0

0.0.2 (2025-12-19)¶

• Feature R distributions, JAX optimizations, initial resampling

  • feat: Richardson extrapolation for exact lmerTest parity

  • feat: numerical gradients for Satterthwaite df

  • feat: prediction uncertainties for LMER

  • feat: initial resampling module (permutation, bootstrap)

  • perf: JAX gradient checkpointing for memory efficiency

  • refactor: consolidate model diagnostics

• Feature JIT, grammar, bootstrap, polars migration

  • perf: JAX JIT optimizations for LMER module

  • perf: build Z matrix directly in sparse COO format

  • perf: keep ZΛ sparse in PLS solver

  • feat: bootstrap enhancements — BCa CIs, semi-parametric, parallelization

  • feat: resampling methods for all model types (permutation, bootstrap, CV)

  • feat: grammar module for declarative resampling inference

  • refactor: migrate data handling to Polars

• Feature simulation, parity testing, consolidation

  • feat: LMER/GLMER bootstrap implementation with parity tests

  • feat: GLM bootstrap and permutation methods

  • feat: complete anova/confint coverage for all models

  • feat: 3+ term random effects support

  • feat: Satterthwaite df for emmeans contrasts

  • feat: boundary tolerance check for singular fits

  • feat: predict(newdata) support

  • docs: user-facing inference guide

• Feature bossanova package extraction, GLM/GLMER maturation

  • feat: extract bossanova as standalone package from nltools

  • feat: rename emmeans API to mee (marginal estimated effects)

  • feat: theta/variance components bootstrap (which=‘theta’, which=‘all’)

  • feat: unified infer() method for post-hoc inference switching

  • feat: inference=None mode for faster fitting

  • feat: InstEval benchmark suite

  • feat: nAGQ parameter for faster GLMER fitting

  • refactor: separate fit() and infer() for cleaner workflow

  • refactor: standardize naming conventions for 0.1.0 release

0.0.1 (2025-11-25)¶

• Bug Fix achieve lme4 parity — 8-phase refactor

  • fix: optimizer settings and REML deviance computation

  • fix: nested/crossed random effects initialization

  • fix: categorical variable encoding to match R’s lme4

  • fix: BOBYQA optimizer defaults to match lme4/minqa rhobeg

  • fix: Satterthwaite df for diagonal (||) structures

  • refactor: systematic parity analysis with R reference maps

  • test: comprehensive parity testing infrastructure

• Feature nltools port and early stats foundation

  • feat: port statistics, resampling, ISC, and adjacency modules from nltools

  • feat: add regress, FDR, Holm-Bonferroni correction functions

  • test: add initial test suite for core stats functions

  • chore: set up package structure with pyproject.toml build

• Feature clean-slate JAX foundation

  • feat: pure JAX implementation with JIT compilation

  • feat: auto-batching memory management for large datasets

  • feat: dual-formulation SVD ridge regression beating himalaya

  • perf: AOT compilation integration for ridge CV

  • refactor: delete legacy nltools code, start from scratch

• Feature LMER v1 — sparse Cholesky solver, formula parsing

  • feat: initial LMER solver with sparse Cholesky factorization

  • feat: formula parsing for random effects specifications

  • feat: symbolic factorization caching for repeated solves

  • feat: model diagnostics and profile likelihood CIs

  • feat: support complex random effects structures (crossed, nested)

  • refactor: replace jaxopt with optax for optimization

• Feature complete model suite — pre-parity baseline

  • feat: implement LM model with emmeans support

  • feat: implement GLM model with logistic/Poisson families

  • feat: initial GLMER with two-stage optimization

  • feat: Satterthwaite degrees of freedom for LMER

  • feat: ANOVA model comparison for mixed models

  • feat: marginal effects estimation and contrasts

  • feat: uncorrelated random slopes (|| syntax)

  • feat: GLMER Hessian-based vcov for lme4 parity

  • feat: R-style distribution functions (60 functions, 15 families)

  • feat: design-based simulation for power analysis