2026-11-02 : Release 0.2.0
- First release under the paper's effect/heterogeneity names (see
  below); the renamed arguments make this a breaking release for 0.1.0
  callers.
- Since 0.1.0 also: the documentation site and the reproduction of the
  paper's simulation study.

2026-11-02 : Use the paper's names for the two hypothesis kinds
- Breaking: renamed interweaved_hierarchy's magnitude_stake and
  difference_stake to effect_stake and heterogeneity_stake.
- Breaking: Claim.kind is now "effect" or "heterogeneity" (was
  "magnitude" or "difference"), and hypotheses are named E[node] and
  H[node] (was M[node] and D[node]), matching E_S and H_S in Wilson
  (2026), section 3.
- USAGE.md maps the paper's direction notation (E+, E-, H^R, H^L) onto
  the `alternative` argument.

2026-11-02 : Remove the old simulation script
- Deleted simulation.py and simulation_results.png. paper/ now
  reproduces the paper's simulation study in full, with the paper's
  own data-generating process, which the old script did not use.
- Dropped the `simulation` extra, which existed only for its xgboost
  dependency.

2026-11-02 : Documentation site and reproduction of the paper
- Added Sphinx documentation (docs/), deployed to
  https://convexanalytics.com/trickle-down/ by
  .github/workflows/deploy_docs.yml.
- Added paper/, a reconstruction of the paper's simulation study from
  its text, and a guide comparing it with the paper's Tables 2-4.
- scikit-learn, sphinx, pydata-sphinx-theme and myst-parser joined the
  dev extra.

2026-11-02 : Rename the package to trickle_down
- The import package is now trickle_down (was pytrickle), matching the
  trickle-down distribution: `import trickle_down`. There is no
  pytrickle compatibility shim; update imports.
- tests/test_pytrickle.py is now tests/test_trickle_down.py.

2026-11-02 : Relicense under Apache 2.0; publish as trickle-down
- Relicensed from MIT to the Apache License 2.0 (LICENSE.txt), the
  licence used by gamdist and PairMatch. Copies distributed earlier
  under MIT remain available under MIT.
- The distribution is now named trickle-down, since pytrickle on PyPI
  belongs to an unrelated project. The import name is unchanged.
- Added a publish workflow: pushing a v* tag builds with uv and
  uploads to PyPI by trusted publishing.
- Rewrote README.md around the paper, with installation, an example
  and a citation.

2026-11-02 : Move to uv, hatchling and GitHub Actions
- Replaced poetry with uv. pyproject.toml now uses PEP 621 metadata
  with a hatchling build backend; uv.lock replaces poetry.lock.
- scikit-learn and xgboost, needed only by simulation.py, moved to a
  `simulation` extra; `dev` holds pytest, mypy, ruff and type stubs.
- Replaced black with `ruff format`, and adopted the ruff rule set
  used in gamdist (E, F, W, I, B, UP, SIM, RUF). zip() calls over
  aligned sequences now pass strict=True.
- Tightened mypy: our own code must be fully annotated.
- Moved pytest and coverage configuration from tox.ini into
  pyproject.toml. Warnings are now errors rather than suppressed.
- Renamed test/ to tests/.
- Added a CI workflow running ruff, mypy and pytest (coverage at
  least 80%) on Python 3.11 and 3.12.

2026-11-02 : Uplift hierarchy, usage guide, and import fixes
- Added the interweaved magnitude/difference hypothesis architecture
  (uplift.py), one-sided and contrast tests, confidence bounds by test
  inversion, and a reference implementation of the graphical approach
  (graphical.py), following Wilson (2026), arXiv:2610.37966.
- Hierarchy.test_hypotheses re-tests every node on each call.
  Previously a node already marked tested kept its verdict, so
  re-running at a new fwer left verdicts that disagreed with the
  levels recorded beside them.
- Hierarchy raises ValueError unless 0 < fwer < 1.
- Packaged USAGE.md, returned by pytrickle.usage().
