BEAST 3 (release 2.8.0, due 3 Dec 2026) is a rewrite of BEAST 2's core. Before it is released we check that it gives the same answers as BEAST 2.7.9 on two kinds of analysis:
The example XMLs. Every example shipped with BEAST 2 is converted to BEAST 3 and both are run, which tests the XML-to-MCMC path model by model. The unmodified BEAST 2 XMLs are also loaded in BEAST 3, as a backward-compatibility check.
Analyses built by hand in BEAUti. 30 analyses are built in BEAUti 2.7.9 and BEAUti 3 by five testers, covering every model option in the templates. This tests what users actually do.
Each analysis runs twice in each version (seeds 1 and 2). Every logged parameter is compared: a z test on the means (flag at z ≥ 3), overlap of the 95% HPD intervals, ESS ≥ 200, and clade probabilities. A flag that can be traced to a known, harmless difference, such as a constant offset in the log prior, counts as passed, explained.
Results are collated on Wed 11 Nov and signed off on Fri 13 Nov.
Each example XML shipped with BEAST 2 v2.7.8 is converted to BEAST 3, both are run with seeds 1 and 2, and the posteriors are compared marginal by marginal. 42 examples are tested; the 2 *BEAST 1 examples are excluded.
1 status line(s) out of step with the numbers.
Each combination is built by hand in BEAUti 2.7.9 and BEAUti 3, run with seeds 1 and 2 in both versions, and compared marginal by marginal. Together the 30 cover every substitution, frequency, clock, tree-prior and starting-tree option in the BEAST 3 templates.
1 status line(s) out of step with the numbers.
The open issues and pull requests that stand between master and the release, sorted for review.