v0.5.3

Trust-focused release: data-quality guard, binary design fixes and pandas 3 support.

New

• Sample Ratio Mismatch (SRM) check — Tester now verifies that the observed group sizes match the expected split (chi-square at the strict 0.0005 level) before evaluating results and warns when the assignment looks broken. Custom splits via srm_expected_ratios={"A": 0.9, "B": 0.1}, disable with check_srm=False. Also standalone: ambrosia.tools.srm.check_srm (pandas & Spark).

Fixed

• Binary experiment design with Bayesian intervals: bayes_beta conjugate prior parameters now reach the size/effect/power computation instead of raising TypeError (thanks to Artem Vasin, #45).
• One-sided alternatives in binary design: alternative="greater"/"less" works in the power table and is honored by the standalone design_binary_* functions (previously silently two-sided).
• Colliding keyword arguments (e.g. confidence_level) raise a clear ValueError instead of silently overriding design parameters; unsupported groups_ratio with the binary method now warns.
• The standalone test() works without an explicit first_type_errors (applies the documented 0.05).

Dependencies

• pandas 3.x is now supported (>=1.5.0, <4.0.0), verified from pandas 1.5.3 up to 3.x; pytest 9 / pytest-cov 7 in dev; security refresh of locked dependencies (pillow 12.2, Pygments 2.20, pytest 9.0.3).

See the CHANGELOG for full details.

---

Note: superseded by 0.5.3.post1 — the SRM check became opt-in (check_srm=True or srm_expected_ratios to enable). Unpinned installs resolve to the post release automatically.