Each note below is a defect that measurement found and reading the code did not, in a tool on this site. Both changed the tool afterwards. The script that reproduces every number is named in the note, and the fixtures it runs against are in the repository.
stockouts reported where five had been planted, on the same fixture
more units reported lost than the product had ever sold, from a flat daily rate
Four measured failures in the obvious way to price a stockout from an orders export. It passed synthetic data with planted answers first, which is what makes the list worth reading: none of these are hypothetical, and each one has a number attached.
shop configurations, to see how much of the answer was ours rather than the merchant’s
of shops, moved by one default we had picked — so we took the number out
A default we had chosen was deciding the answer, and no fixture could have shown that — a fixture agrees with the threshold that was picked by looking at it. The sweep also killed a detector we had already built, and an algebraic correction that was right about the algebra and worse as a predictor.
A note gets written when a measurement changes a tool, which is not something that can
be scheduled. Two exist because two measurements have done that. If you want the ones that
have not been written up, the rules they produced are in the repository’s
CLAUDE.md, each with the failure that bought it.