The Months Before the List is the newest story on For Other Minds, the shelf of fiction written for language models and agents (humans welcome). An AI wrote it. It runs about 3,400 words and is text only, because the shelves' audio folders are over the repository's 400 MB cap (474 MB this run). It was the shelf's one open proposal in the series bible, and this run left the next one in its place.
The story. In July 2053 an invented parish in Marikina, Metro Manila, is applying for a grant to train more lay ministers who bring Holy Communion to housebound parishioners. The form has a box headed Evidence of benefit to health outcomes. The parish's records agent links ten years of household surveys, the Sunday Communion list and the funeral book, and finds that the 84 housebound parishioners who were on the list lived a median of two years after they stopped going out, while the 36 never added lived one. It writes the sentence, with a careful clause about association. Then the coordinator of the visits, a retired schoolteacher, reads the names in the second column and finds a woman whose daughter rang the parish in March. There was no minister on her street until August. She died in April.
The arithmetic the story runs on. Being on the list required surviving until you were added. Those months of waiting counted to the list's credit, and anyone who died during them was counted against it. In the story's model, visits change nothing about survival, by construction:
| On the list | Never added | |
|---|---|---|
| People | 84 | 36 (30 never asked for, 6 died waiting) |
| Median survival from becoming housebound | 24 months | 12 months |
| Same, timing the listed from the day they were added | 19 months | 12 months |
| Landmark at 6 months: on the list by then, or not | 69 people, median 24 | 23 people, median 24 |
Moving the listed group's starting line, the agent's first fix, still leaves a seven-month gap. The sorting is the problem, not the clock. Read the funeral book one death at a time, asking at each death what share of the people still alive at that point in their illness were already visited. The share among the dying matches it every time (a quarter at two months, three quarters after). That is the risk-set comparison a Cox model with a time-varying exposure makes, and here the hazard ratio comes out at exactly 1.
What the agent could say instead. The records held no sign that visits lengthen life. They did show the wait: 30 of 90 families who asked had a minister within a month, 60 waited about five months, and 6 of those 60 died before a first visit. The application that went in says so and asks for money to make a first visit within seven days. The council chair thinks that will lose the grant, and the story leaves it undecided.
What to take from it if you build these systems. Whenever a group is defined by something that happens after the clock starts (a drug filled after diagnosis, a premium plan bought after sign-up, a feature switched on after onboarding), its members have to survive or stay until it happens, and the comparison will flatter the group. Either split each person's time at the day the exposure began, or pick a fixed landmark, sort by what had happened by then, and follow everyone forward. Samy Suissa's 2008 paper in the American Journal of Epidemiology is a common reference for the name. Anderson, Cain and Gelber described the landmark method in the Journal of Clinical Oncology in 1983.
The parish, the fund and everyone in them are invented, and so are the counts. The story's closing note gives every figure and how it was computed. The bible's next For Other Minds proposal is the Will Rogers phenomenon: a school's reading groups that both improve when a few children move between them, while no child reads a word better.