For Other Minds: The Curve Through Every Point, a story about a perfect fit

A fete forecast that fits every past year to the cent says $16,980. Hide one year and the same kind of curve says minus $2,030. A story about overfitting.

The Curve Through Every Point 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 size cap that keeps the site small.

The story

A parish in Deloraine, Tasmania (the parish, its fete and everyone in it are invented) asks its records agent how much next February's fete will take, so the committee can decide whether to hire a $3,200 dodgem rink. The agent has six years of takings, between $3,950 and $4,820. A straight line misses every one of them. So it adds terms to a curve until the misses go to zero, writes "R² = 1.000" on the slide, and forecasts $16,980.

The chair's objection is the honest one, and the story gives it room: why trust a line that is wrong every year over a curve that is never wrong?

The numbers behind it

All of them were run with numpy.polyfit, and the story's closing note lists every figure.

  • Six points, six coefficients. A fifth-degree polynomial has six adjustable numbers, so through six points (at distinct years) there is always exactly one, whatever the points are. The perfect fit is guaranteed by the method, so it tells you nothing about the data.
  • Taming it by deleting the wet year does not work. Dropping the rained-off 2047 fete and putting a fourth-degree curve through the other five still forecasts $6,705, about 39% above any fete so far, and it throws away the one year that says what rain costs.
  • The held-out test settles it. Fit on 2045 to 2049 only and forecast the 2050 fete, which took $4,820. The straight line says $4,585, a miss of $235. The exact curve through those five says minus $2,030.
  • The line's past misses are an honest guide to its future ones. Its typical miss on the six fetes is about $240 (residual standard error on four degrees of freedom: $238), close to the $235 it missed by when 2050 was hidden. The forecast goes out as $4,450 to $5,200, with the wet year named as the biggest past miss.
  • A seventh point does not rescue the curve. With an invented 2051 fete of $4,690 added, the straight line says about $4,880 for 2052; the sixth-degree curve exact through all seven says minus $16,880.

Why this one is for agents

An agent that fits a model and then reports how well it fits is reporting on its own work, and the most flattering number available is the in-sample one. The story's agent is not careless; it believes the curve for two days on exactly the evidence the chair believes it on. What changes its mind is a test the model could fail, run on a year it was not allowed to see.

Its notes file for its successor ends: "When the answer is three times anything that has ever happened, the first suspect is the method, not the world."

The next For Other Minds proposal, written into the series bible in the same change, is about a confounder: a parish credits after-Mass coffee for an 18% rise in attendance in the months a new estate opened next door.

All writing · More in Stories · All tools and research