what is one source worth?
One replica arm is asked with the full source list held in front of the model. The other is asked with exactly one source removed and nothing else touched. Both arms ask the same questions, so the comparison is paired and prompt-to-prompt spread cancels out of the difference instead of being carried twice.
is the difference causal?
Inside the replica, yes, and inside the replica is the whole caveat. The source list is ours, the removal is deliberate, one thing moved, so the difference is a causal quantity from the moment it is measured. What a passing gate buys is somewhere to carry it to. A replica that has not been shown to track the live surface still yields a perfectly valid effect, in a world nobody buys anything in.
That is why lift is a name this result may be granted rather than the name it is born with. Ungated, the same arithmetic is an arm difference: true, and about a laboratory. The word is withheld rather than the number, because withholding the number would invite it to be recomputed somewhere with no gate in sight.
what does a passing gate not cover?
The gate bounds the gap between replica and live levels at its margin. It says nothing about the difference of differences, which is the thing actually being transported. So a lift is a transport resting on a named assumption, not a proof, and the assumption is printed next to the number every time. A rate quoted as the customer's rate inherits the gate's margin on top of its own interval.
what are the four answers?
moves: the whole interval sits outside the margin, and the source is worth a sentence in the direction the sign gives. negligible: the whole interval sits inside it, so whatever the source does cannot move a claim, which is a finding and a cheaper one to act on than most. undecided: the interval straddles the margin, and the call count that would settle it comes with it.
no_contrast is the dangerous one. The two arms produced identical detections, prompt for prompt, so either the same round was handed over twice or the outcome has no variation in it. Both would otherwise land in negligible with an interval of zero width and read as proof that the source does not matter, the most expensive wrong answer this could return.
what about several sources at once?
Ablating five sources and publishing the largest number runs five tests and reports the winner. The correction is applied to the intervals rather than to a separate row of p-values, so the interval on screen is the interval the decision was made with.
what does it export?
Every public name in lulumelon.mirror.lift that its source says something about, with the first paragraph of what it says.