story-shapes

What happens to the hero, two hundred words at a time.

Kurt Vonnegut drew stories on a blackboard: good fortune up, ill fortune down, beginning to end. Here a small language model, Kev-4B, reads 29 classics passage by passage and answers one question each time: how are things going for the protagonist?

Kev's reading, smoothed 95 % band: where the curve could be shape written down before reading G good fortune · I ill fortune · B beginning · E end

Did it find the shapes we expected?

Pre-registered test

Before reading a single book we wrote down each one's expected shape and what counts as a hit: the nearest shape is the expected one, and the band is narrow enough to rule out a flat line.

Post hoc, not pre-registered

Drop the second bar and count a hit whenever the nearest shape is the expected one. Kev mostly points the right way; a few hundred passages are not enough to say so with confidence. The pre-registered result is the one that stands.

Candide and Hamlet have no agreed shape and are shown as controls. Full scoring

All 29 books, grouped by the shape we expected

hit: nearest shape, band rules out flat nearest shape, band too wide read as another shape

How a curve is made

01 Cut

Each book comes from Project Gutenberg and is cut into passages of about 200 words. The model never sees the title, only the protagonist's name and the passage.

02 Ask

Kev-4B answers on five outcomes, from "they lose what matters most" to "they get what they most wanted". Outcomes, not feelings: gloomy words around good news no longer fool it.

03 Smooth

The answers are smoothed into one line with a 95 % band that widens where passages disagree or the model hesitates. The line stops where the protagonist dies for good.

15 books were read on a Mac (MLX, bf16) and 14 on a Kaggle T4 (torch, fp16). Daisy Miller, read on both, agrees at r = 0.9999.