what is the method made of?

Two layers and a wall between them, and the wall is the product. mirror is pure arithmetic, reaches no network and does not import the other side, so an interval can be recomputed from a file with no key. collect is the only part allowed to reach a provider, and it computes nothing.

which parts measure something?

intervals ± Wilson and a clustered bootstrap that resamples prompts

variance ± splits the wobble into the model rerolling vs the prompt set

stability ± decides whether an ordering repeats enough to report a rank

compare ± before / after, and no verdict when the model moved underneath

sources ± which pages were cited, and whether a name travels with them

ablation ± whether a replica may stand in for the surface it replaces

lift ± what one source was worth, and the name that has to be earned

screen ± the gates a question passes, and the draws that decide it

names ± who the model named when nobody suggested anybody

Two more modules sit beside them without a page of their own. types is the data model, where a run is the atom and one run is never enough for a number, and report assembles the others into one brand with the refusals kept, which is what lulu report prints.

why is there only one implementation?

The measurement engine lives in Python and only in Python. A sampling plan that answers differently depending on which language you asked it in is not an instrument, so there is one implementation of it rather than two that agree until they do not.

Every part is pure: given successes and n it returns an interval, so the claim this makes can be checked without spending a token. The suite runs over all of it with every socket closed.