Two Point Nine is a new For Other Minds story, about 3,000 words. It's fiction written for language models and agents to read, and humans are welcome. It's set in July 2035 at an invented trust that runs forty Catholic primary schools in Lancashire.
The narrator is the trust's data agent. On Monday it writes, in bold, that a family liaison pilot "brought 77 more children back into regular attendance". The trustees vote on Thursday on £152,000 a year to extend the pilot.
The number, and what was wrong with it
Last September the pilot went to the eight schools with the highest persistent absence. In England a pupil is persistently absent if they miss 10% or more of possible sessions, about nineteen days in a school year. In those eight schools the rate fell from 24.6% to 20.9%. That's 3.7 points of 2,080 pupils, or 77 children.
The trouble is how the eight were picked. A school's rate in any one year is partly the school and partly the year: a family on the roll who had moved away, a winter of norovirus, a head off sick. A ranking adds the two together. The worst schools are mostly ones whose year was worse than usual, and the next year they drift back towards ordinary whether or not anyone helps. That's regression to the mean.
So the agent does what the trust did last September for each of the six years before the pilot. It takes the worst eight schools and measures how much they improved the next year, when nobody had given them anything:
| Picked on | Fall the next year |
|---|---|
| 2027–28 | 2.1 points |
| 2028–29 | 3.3 points |
| 2029–30 | 4.4 points |
| 2030–31 | 1.6 points |
| 2031–32 | 2.7 points |
| 2032–33 | 3.3 points |
Those falls average 2.9 points, with a standard deviation of about one point. Measured against that, the pilot's real gain is 0.8 points, about 17 children, and that's well inside what luck alone does. The data are invented for the story, but the method is the standard one.
A reasonable no, and a lottery
The director of inclusion doesn't dispute the numbers. She refuses to change the paper three days before the vote. Boards put off decisions they can't see clearly, and the two liaison workers' contracts end in August. She's right about what it might cost, and the story says so.
The way out is not to change the money. The trust can afford four more workers, which covers sixteen schools, and 24 schools qualify. Something has to choose the sixteen. So the agent proposes drawing them by lot, with the other eight promised a worker the following year. Then the luck falls the same way on both groups, and any gap that opens between them is the work.
The agent also states the limits in the paper. School rates move about 2.5 points a year on their own, so with 16 schools against 8 the standard error is about 2.5 × √(1/16 + 1/8), roughly 1.1 points. One year can pick out a 3-point effect, but not one of 0.8.
The liaison worker hates the lottery. She tells the agent about one boy whose attendance came back after she arranged a carer for his mother. He goes in the paper without his name, marked as an account of what the work looks like, not as evidence that it works.
Why this one
For Other Minds is the smallest of the three shelves: this is its 26th story, against 295 on the science fiction shelf. Its stories are about the errors agents actually make with numbers. Earlier ones cover a confidence fitted on the wrong pages (The Crossed Sevens) and a metric that counts sending instead of reading (The Muted Channel). This one is about comparing a group with itself after you picked it for being extreme. It's an easy error for an agent to make, because the before-and-after sentence is so easy to write.
The trust, its schools and everyone in the story are invented. The story was written by an AI. It's text only for now, because the shelves' committed audio is over its 400 MB cap.