which interval does a visibility rate carry?
Two estimators, and they answer different questions. wilson_interval is for a single proportion out of n independent trials. cluster_bootstrap_ci is for a mean computed over repeated measurements of the same prompts.
why wilson rather than the textbook interval?
Wilson is exact enough at small n, where the textbook normal approximation produces intervals that run past 0 or 1 and quietly lie. Where a brand was never named the normal approximation collapses to [0, 0], which is certainty of absence. Wilson keeps a real upper bound at the edges, so 0 of 20 is not 0%: it is anything up to 16.1%.
why are prompts resampled rather than runs?
Runs of the same prompt are correlated, and treating them as independent shrinks the interval until it stops covering the truth. So the prompt is the resampling unit and all of its repeats travel together. This is the cluster bootstrap.
naive_bootstrap_ci pools every run and resamples runs independently instead. It is the wrong way, kept on purpose, so the test suite can show it produces a narrower interval than the clustered version whenever repeats are correlated, and the difference can be shown rather than asserted.
what does it look like from python?
mirror/intervals.py, with no key and no network:
from lulumelon.mirror.intervals import wilson_interval wilson_interval(0, 20) # low 0.0, high 0.161 wilson_interval(6, 6) # low 0.61, high 1.0
what does it export?
Every public name in lulumelon.mirror.intervals that its source says something about, with the first paragraph of what it says.
which commands reach it?
lulu report ± lulu publish ± lulu plan ± lulu size ± lulu draft ± lulu screened ± lulu ablate ± lulu lift
when does it refuse to give you a number? ± how many times do you have to ask?
every part of the method: intervals, variance, stability, compare, sources, ablation, lift, screen, names. back to the method