One read covers a group, not a twin
Before a question is asked, the twins of a place are sorted into cells: groups of twins who are alike on the things most likely to move an answer. The model reads one twin from each cell — the one nearest the middle of its group — and every twin in that cell carries the answer that read produced.
So a cell holding hundreds of twins is one observation of the place, not hundreds. Counting it as hundreds would make the interval look far tighter than the evidence behind it.
Counting what was actually observed
The interval is built on the effective sample over the cells that answered: how many fully independent reads the result is worth once each cell’s weight is taken into account. A cell standing for a large share of the place counts for more of the estimate, and the more unevenly weight is spread across cells, the smaller the effective sample becomes.
The number of twins never enters that count. That is why the interval on a reading is wider than a reader used to sample sizes might expect from the number of twins — the width is the price of the reads that were not taken.
A cell with no answer adds nothing
Sometimes a cell’s read does not come back. That cell still gets an estimate, borrowed from the cells most like it, so the place’s make-up is kept whole. But it adds nothing to the effective sample, because nothing about it was observed. The count of groups that went unanswered is kept and reported beside the interval rather than hidden inside it.
Groups within the place are priced the same way
A breakdown row — renters, an age band, a neighbourhood — gets its interval from the cells its twins sit in, each counted by how much of that group it holds. A group spread thinly over many cells is priced on many light units, and uneven spread shrinks its effective sample just as it does for the whole place.
A group reached by too few answered cells keeps its share and its counts but gets no interval, and the reading says why: with so few reads, a width would come from the formula, not from anything observed about the group.
What the interval does not cover
The interval prices how much independent information a reading holds. It does not price whether the model read a group well: a reading whose cells were read badly can come out narrow and still be off. And when twins of two groups share a cell, they share its one read, so the reading cannot see a difference between them there. Each breakdown row is checked for this, and the check is written into a reading’s downloaded breakdowns beside the row, so a difference between groups that share cells is not taken for one the reading observed.
Both limits are reasons to read a result as a contrast, with its interval, rather than as a level.