Organization beyond project, tested with the project as the unit¶

The client-level permutation treated clients as exchangeable, and clients share their project. The honest test leaves each client's own project out of every class mean, so an organization has to be recognised from its other projects, and permutes organization labels across projects, not clients (attribution.group_permutation_p, exact over all distinct relabelings when few enough). On the 34 clients of job 148513:

Suggestive and not significant, and eight projects cannot give much less than this. No single content type has two projects on both sides, so the language-controlled version cannot run here. The AOSP and Qt C++ cell (seven and six projects, 1,716 relabelings) is built for exactly this test.

The review of the packing fix caught a confound in the fix itself. First-fit decreasing opens clients in order of change size, so keeping the first limit kept the clients built around each project's largest changes: one change and one reviewer each for a project with big changes, twenty small changes for another. The attack could have read the packing instead of the source. Clients now take at most a quarter of their examples from any one change, and the limit kept are a seeded random draw from all full clients. The executed run predates both packers' flaws only in count; the next run is the first under the corrected one.