can a replica stand in for the live engine?
A claim about what one source does to a brand's rate is causal, and nothing observational can support it: sources and appearance move together for a dozen reasons. The only way to get at it is to control the sources, which means asking the same model the same questions with the source list supplied as context. That is a replica, and a replica is a laboratory instrument.
why is the replica tested first?
Before anything is concluded from one, it has to be shown to behave like the surface it stands in for. If it does not, the whole line of work is a laboratory toy, and finding that out costs one round rather than three months.
why is overlap the wrong test?
Two rates whose intervals overlap can be genuinely alike, or they can be two noisy measurements that were never asked to disagree. A comparison with too few calls overlaps with everything, so a gate built on overlap rewards not looking. Reading a non-significant difference as sameness is the same error wearing the opposite hat: absence of evidence arriving as evidence of absence.
So the gate is an equivalence test. It asks whether the difference between replica and surface is provably smaller than a margin named in advance.
what are the three answers?
stands_in: the whole interval sits inside the margin, so reasoning from the replica is reasoning at the precision the product quotes its own findings at. differs: the whole interval sits outside it, and the replica is not the surface. undecided: the interval straddles the margin. This design cannot tell the two apart, which is a fact about the design and not about the replica, and it comes with the number of calls that would settle it, computed rather than guessed.
who picks the margin?
Nobody, by taste. It is the half-width the measurement is planned at. A tighter margin would reject a replica for a difference nothing downstream can see, and a looser one would accept a replica capable of moving a published figure. The call count is not a taste either: it comes from the same sampling model the rest of the library sizes rounds with, on the measured split, and carries the factor of the square root of two that a paired comparison costs.
what does it export?
Every public name in lulumelon.mirror.ablation that its source says something about, with the first paragraph of what it says.