What’s new#

A digestible summary of recent releases — for the full per-line detail, see CHANGELOG.md.

0.29.0 — best-practice infrastructure (May 2026)#

Theme: forward-compatibility + observability + perf-regression detection. No new user-facing API surface; better support for the existing one.

Highlights#

  • @deprecated decorator for forward-compat planning. Mark public symbols with a string deadline: @deprecated("0.31.0", reason="…", use_instead="…"). Validated at decoration time; emits DeprecationWarning on every call. Policy in docs/DEPRECATION.md: minimum two-minor-versions warning before removal.

  • Structured library logging. eval_toolkit.{harness,leakage, bootstrap,loaders} each emit DEBUG records (silent by default via NullHandler at the root). Filter granularly: logging.getLogger("eval_toolkit.harness").setLevel(logging.DEBUG).

  • pytest-benchmark on math kernels. Weekly cron in nightly-benchmarks.yml records per-kernel timings; catches perf regressions before release. Excluded from PR CI to keep iteration fast.

  • docs/RELEASING.md runbook. Concrete release recipe with every gotcha + recovery from the v0.27.x and v0.28.x cycles.

  • .editorconfig + CODEOWNERS. Repository hygiene.

Migration#

No migration needed — v0.29.0 is fully backward-compatible with v0.28.x. Logging is opt-in; deprecation decorator is internal tooling, not yet applied to any public symbols.

0.28.1 — security patch (May 2026)#

Theme: supply-chain security signal.

  • CodeQL workflow (push/PR/weekly cron) — static security analysis with the security-extended query suite. Findings populate the Security → Code scanning tab.

  • pip-audit gate in test-base-install — fails PR CI on any known CVE in runtime deps (numpy / scipy / scikit-learn / jsonschema). Dev-extras vulns are surfaced via Dependabot.

No source-code behavior change.

0.28.0 — temporalcv cross-pollination bundle (May 2026)#

Theme: cross-pollinate from sibling temporalcv + hosted docs.

  • PurgedKFoldSplitter (from eval_toolkit import PurgedKFoldSplitter) — time-aware k-fold with explicit purge gap + post-test embargo, preventing label-window leakage when labels have a forward horizon. Per López de Prado 2018 Ch. 7.

  • compute_label_overlap(t_train, t_test, horizon) — standalone helper for auditing arbitrary train/test label overlap.

  • Nightly Monte Carlo bootstrap CI calibration — validates empirical coverage of nominal 95% CIs across pr_auc / roc_auc / imbalanced cases. Catches “math is wrong but self-consistent” bugs that goldens can’t.

  • 6-example documentation gallery — minimal sybil-validated Markdown examples covering metrics + bootstrap, evaluate harness, calibration, leakage detection, claims/gates, paired comparison.

  • Hosted mkdocs-material docs site at brandon-behring.github.io/eval-toolkit. Auto-generated API reference from NumPy-style docstrings via mkdocstrings. Full LaTeX + TikZ rendering via MathJax v3 + tikzjax.

  • Public-repo polish: SECURITY.md disclosure policy, CITATION.cff academic citation, 5 README badges.

0.27.2 — base-install fix (May 2026)#

Theme: silent bug fix — pip install eval-toolkit was broken for users who didn’t also install [dataframe].

  • Fixed import pandas as pd at module top level in four modules (harness, loaders, leakage, splits) so base install works for the headline from eval_toolkit import evaluate path.

  • Added a regression-guard CI job (test-base-install) that installs without extras and verifies imports.

0.27.1 — first PyPI release (May 2026)#

Theme: eval-toolkit goes public.

  • First release on PyPI: pip install eval-toolkit.

  • Auto-publish on v* tag via PyPI Trusted Publishing (OIDC).

  • Bumped from v0.27.0 because the internal milestone tag of that version pre-dated the PyPI publishing infrastructure.

Prior releases#

For the full v0.26 → v0.27 → v0.28 progression and individual release CHANGELOG entries, see CHANGELOG.md.