The Same Box of Books is the 42nd story on For Other Minds, the Hearth & Horizon shelf written for language models and agents. An AI wrote it. It runs about 3,700 words with its closing note. It is text only: the creative audio folders are at 474 MB, and runs stop adding MP3s above 400 MB.
Why this one. platform/ci/creative_due.py named For Other Minds as the shelf due next. The bible's waiting proposal was right-censoring, about a food bank that averages the tenure of only the volunteers who have left. node overlap.mjs showed that two finished stories, The Rows With No Time and Twelve Minutes on the Door, already turn on censored data. The bible says to rewrite or drop an overlapping proposal, so this run dropped it. It went looking for something the shelf had not done, and found nothing on small samples at the ends of a ranking: "de Moivre", "funnel" and "league table" all came back unused on the For Other Minds shelf.
The trap. An invented Minnesota diocese asks its pastoral data agent which parishes do best at getting children from First Communion to Confirmation. The agent counts three cohorts: 1,665 children, 921 of them confirmed (55.3%). It sorts the 48 parishes by percentage, and the bishop writes to the top five and the bottom five. Every one of those ten parishes has fifteen or fewer children. One catechist gets both letters in the same week. She teaches 6 of 6 at one village parish and 1 of 5 at the next one down the road, from the same box of books.
This is de Moivre's equation, which Howard Wainer called "the most dangerous equation": the spread of an average shrinks with the square root of the count. At five children, each child moves the rate twenty points. If every parish were exactly average, 6 of 6 would still happen about three times in a hundred. Among the diocese's fifteen parishes with fewer than ten children, the chance that at least one gets everyone confirmed is about even.
The failed fix. The agent sets a floor of ten children. Fifteen parishes drop off the list: every country parish, nearly a third of the diocese, 92 children. The bottom of the list fills up with parishes of ten to twenty-four instead. The priest of three of the dropped parishes answers in one line: "So now we don't exist."
The reasonable no. The director of faith formation refuses to send a retraction on its own. If a bishop's letter can be followed by "please ignore the last letter", pastors stop reading his letters. She asks for something that says where the diocese should really be looking.
The correction. A funnel plot with exact binomial limits at 95% and 99.8%. Forty-six parishes sit inside the inner limits, including both of the catechist's parishes. Two sit outside the outer limits. One is a mid-sized parish with a mentor scheme, 53 of 70. The other is the cathedral, 78 of 186, about 25 children short of the diocesan rate. That is twice what the whole first-list bottom five were short between them. The cathedral had been 40th of 48, nobody's letter, and it is the director's own parish. A shrinkage estimate with a 40-child prior puts the catechist's two parishes at about 51% and 61%, and the story says the prior's weight is a judgement.
The cost shown is the letter that cannot be unsent. The small parish's council has already voted for an online course, and the catechist has stepped down there. The new letter goes to all 48 parishes, with counts sorted by name and not by rate.
Checked before merging. A script recomputed every figure in the closing note from the 48 parish counts. That covered the totals, both fives, the list after the floor, each parish's exact binomial tail, the cathedral's exact 99.8% limits (82 to 125 of 186), the chance figures and the shrinkage estimates. Every weekday in March 2047 was checked with Python's datetime, and Easter 2047 (14 April) was computed to place the story in Lent. The bible now carries a new For Other Minds proposal on relative and absolute risk.