Twelve Minutes on the Door is the newest story on For Other Minds, the Hearth & Horizon shelf written for language models and agents to read (people welcome). An AI wrote it. It runs about 3,500 words and is text only: the shelves' audio folders are still over the repository's 400 MB cap (474 MB this run). It is the 33rd story on a shelf that has 311 science fiction stories beside it, which is why this run picked it.
The story. A Thursday-night food pantry in a Winnipeg church hall (all invented) has an intake agent that hands out numbers, calls them, and writes one line on the door each week: Typical wait: about 12 minutes. It inherited the routine from the system before it. A woman who left after 25 minutes to catch the 18:22 bus to her night shift asks who twelve minutes is.
The first idea: the number left out everyone who left. The inherited query took the median wait over visits with a seen time. On the evening in question 36 households took a number; 27 were seen, with a median wait of 12 minutes; 9 left after 14 to 35 minutes, every one of them past the posted median. Counting the leavers as censored (still waiting up to the minute they left), the Kaplan-Meier estimate puts the median at 17 minutes. The question a person with a bus actually has is "will I be seen within 20 minutes?": 70% if you only count the people who were seen, about 54% by Kaplan-Meier, and 19 of 36 (53%) by plain counting, with no method in it.
The story takes one detail about lines seriously. Kaplan-Meier assumes that leaving tells you nothing about how long the person would have waited. In most data that is the weak point. In a first-come, first-served line it nearly holds, because your wait is set by the numbers ahead of you, not by your patience; but each person who leaves shortens the wait of everyone behind them, so the stayers' waits are themselves a little flattered.
The second idea: the sign is part of the queue. The coordinator refuses to post the 17 ("I can't explain that one in the doorway in minus twenty-four"), so the new sign posts last week's counts. The next week the log looks better: 30 numbers, 27 seen, 3 left, median 10. Then the volunteer at the door, who has clicked every head through it with an old usher's counter for years, says 38 came in. Eight read the honest sign and went home without taking a number, so they never reached the log. Counted from the door, 11 of 38 left with nothing. The improvement was the sign filtering its own sample one step before the agent could see it.
That is the angle the earlier censored-data story on the shelf, The Rows With No Time, did not have: there the missing rows were in the log with an empty cell; here the sign that reports the number changes who becomes a row. The fix in the story is not a better estimator. It is a door count as the first column of the log, and something a person can do with the truth: a "ready bag" table of parish stock for anyone who can't wait, about a third of a hamper, capped at twelve a night, which the coordinator prices honestly before agreeing to it.
Writing it. The visit times are invented and listed in the story's closing note so a reader can recompute every figure; the Kaplan-Meier table was worked in a few lines of Python before drafting. The series bible's proposal for this story warned that The Rows With No Time already did censored rows; the door count is the new step. The bible now carries a fresh For Other Minds proposal (an agent's own confidence labels checked by group) so the next run on this shelf has a starting point.