does the answer agree with itself across repeats?
A visibility number can be perfectly reproducible in its average and still be built on answers that share almost nothing with each other. So stability is its own metric, not a footnote on the average.
what does low stability mean?
If it is low, every ranking built on top of a single run is a coin flip, and this is what says so with a number. It is why no rank is printed where the ordering does not repeat, and why rank is kept separate from rate: appearing twice at rank 1 is a different position from appearing nine times at rank 5.
which kinds of agreement does it measure?
Set overlap between two answers, ignoring order. Rank-biased overlap, which does not treat named first and named last as the same event, so disagreement about who leads costs more than disagreement about who trails. How often the first name repeats the most common first name, and how often a whole ordering repeats the most common ordering exactly. Each is averaged over every pair of repeats of one prompt, and the report reads them per prompt.
what does it export?
Every public name in lulumelon.mirror.stability that its source says something about, with the first paragraph of what it says.
which commands reach it?
when does it refuse to give you a number?
every part of the method: intervals, variance, stability, compare, sources, ablation, lift, screen, names. back to the method