Suspense Atlascomputational literary science

Aggregate patterns · RQ3

Are suspense trajectories genre-specific?

Mean curves by genre, era, length and author; a permutation test on curve dispersion; clustering of curve shapes against a shuffled null. Primary measure: LLM suspense (Qwen3-8B).

01

By genre

0%25%50%75%100%-1-0.500.51narrative progressz (within story), smoothed
all storiesadventure (n=11)comic (n=11)detective (n=10)ghost (n=13)horror (n=18)literary (n=15)speculative (n=14)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

0.006permutation p, between-genre share of curve variance (2,000 shuffles)
9.7%share of curve variance between genres (observed)
0.027Kruskal–Wallis p, peak position by genre
0.002Kruskal–Wallis p, late−early level by genre
Median shape descriptors by genre (LLM suspense (Qwen3-8B), z-scored curves). Peak position is the fraction of the story at which the smoothed curve is highest.
genrenpeak positionlate − earlypeaks > 0.5 sdfinal level
adventure110.850.512.000.48
comic110.590.333.000.12
detective100.790.923.000.30
ghost130.921.193.000.85
horror180.951.112.001.00
literary150.870.812.000.42
speculative140.820.783.000.56

02

By era and by length

Publication era is confounded with genre and author in a corpus this size; treat differences as descriptive. Length bands test whether the 0→1 normalization hides a real dependence on absolute length.

Era

0%25%50%75%100%-1-0.500.51narrative progressz (within story), smoothed
1880–1899 (n=31)1900–1928 (n=39)pre-1880 (n=22)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

Permutation p = 0.696; between-era share of variance 1.8%.

Story length

0%25%50%75%100%-1-0.500.5narrative progressz (within story), smoothed
long (>9k) (n=23)medium (4–9k) (n=41)short (<4k) (n=28)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

Permutation p = 0.237; between-band share 2.6%.

03

By author

Authors with at least three stories in the corpus. Author identity is the strongest nuisance variable for a text-based measure: register, sentence rhythm and vocabulary are all author-level.

0%25%50%75%100%-1012narrative progressz (within story), smoothed
Ambrose Bierce (n=4)Anton Chekhov (tr. Constance Garnett) (n=3)Arthur Conan Doyle (n=9)Edgar Allan Poe (n=9)H. G. Wells (n=9)James Joyce (n=4)Katherine Mansfield (n=3)M. R. James (n=4)Nathaniel Hawthorne (n=3)Saki (H. H. Munro) (n=6)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

04

Clustering curve shapes — against a null

K-means on z-scored curves for k = 2…6. A silhouette is only meaningful next to its null: the same clustering on within-story shuffled curves. Where the observed silhouette does not clear the null’s 95th percentile, the “archetypes” are what k-means finds in noise.

ksilhouettenull meannull 95th pctbeats null?
20.1060.0340.040yes
30.1080.0310.038yes
40.1010.0270.031yes
50.0960.0240.032yes
60.0840.0190.027yes

k = 3 centroids (shown regardless of whether they beat the null)

0%25%50%75%100%-101narrative progressz

Genre mix: comic 5 · literary 4 · speculative 4 · horror 3 · adventure 3 · detective 2

05

Same question, other measures

The genre picture should not depend on one operationalization. Below, the by-genre curves for each alternative measure with its permutation p-value.

Composite lexical proxy · permutation p = 0.817

0%25%50%75%100%-0.500.5narrative progressz (within story), smoothed
all storiesadventure (n=11)comic (n=11)detective (n=10)ghost (n=13)horror (n=18)literary (n=15)speculative (n=14)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

GPT-2 surprisal · permutation p = 0.011

0%25%50%75%100%-0.500.51narrative progressz (within story), smoothed
all storiesadventure (n=11)comic (n=11)detective (n=10)ghost (n=13)horror (n=18)literary (n=15)speculative (n=14)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

Threat lexicon · permutation p = 0.842

0%25%50%75%100%-0.500.5narrative progressz (within story), smoothed
all storiesadventure (n=11)comic (n=11)detective (n=10)ghost (n=13)horror (n=18)literary (n=15)speculative (n=14)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

Uncertainty · permutation p = 0.098

0%25%50%75%100%-0.500.5narrative progressz (within story), smoothed
all storiesadventure (n=11)comic (n=11)detective (n=10)ghost (n=13)horror (n=18)literary (n=15)speculative (n=14)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

LLM intensity · permutation p = 0.371

0%25%50%75%100%-1-0.500.51narrative progressz (within story), smoothed
all storiesadventure (n=11)comic (n=11)detective (n=10)ghost (n=13)horror (n=18)literary (n=15)speculative (n=14)

Select one group to see its 95% bootstrap band over stories. Curves are z-scored within story before averaging, so height differences between stories are removed and only shape remains.

06

Extremes

The earliest and latest peaks in the corpus on the primary measure.