Literature Mutations

Can the genre system of English fiction be recovered from prose alone — and can we measure the rate at which genres form?

This began in 2021 as a course proposal at Columbia (the original proposal is preserved in docs/PROPOSAL.md). It has since been rebuilt into a working pipeline with real results — one positive, one negative, both reported honestly below.

---

TL;DR

• Yes to the first question. An unsupervised network built from raw opening prose recovers the recognizable genre system of English fiction — detective, science fiction, nautical adventure, historical romance, the early English novel, American realism — and each emergent cluster is confirmed by held-out genre labels the model never sees.
• Mostly no to the second. There is no global mutation rate — the apparent acceleration is a corpus-density artifact, and the event count is indistinguishable from shuffled publication dates (null model, z = −0.27). But after controlling for three confounds (density, style drift, author voice), detective fiction stands out as one genuine, datable genre emergence (books of 1878–1926, z ≈ −3.0). Genre birth is real but rare and genre-specific, not a smooth rate.

---

Corpus: canon-first, cross-referenced

Earlier drafts scraped whatever was available (Goodreads shelves; then bulk Project Gutenberg). Both are biased and unrepresentative. Instead the corpus is now defined top-down and cross-referenced:

1. build_canon.py assembles the pre-1929 public-domain English-novel canon by querying many sources — named critic lists (Guardian 100, Modern Library), reference lists (1001 Books), era/genre buckets, and a crowd source (4chan /lit/) — each confirmed by two independent model families (Gemini, grounded in Google Search; and Claude). Every title gets a support score: how many lists and how many models back it. → 440 titles, 159 cross-model confirmed.
2. build_corpus.py then goes and finds each title on Project Gutenberg (a title+author match is a third, independent existence check), pulling ~20k words of real full text. → 345 titles matched, spanning 1660–1928.

No model ever writes the text we analyze. LLMs only enumerate citeable list membership and verifiable facts; the signal is always real authorial prose.

Method

• Text → vectors (semantic_edges.py): TF-IDF over each novel's prose, dropping vocabulary common to >40% of books (shared "novel-ese") so edges reflect distinctive genre vocabulary.
• Vectors → graph (temporal_network.py): each book links to its k nearest neighbors. (A global similarity threshold fails — over long English prose every novel is somewhat similar to every other, giving one blob; k-NN recovers structure robustly.)
• Graph → genres: Louvain community detection. The corpus is also grown year-by-year to produce cumulative snapshots for the temporal analysis.
• Validation: Gutenberg subject labels are held out and never used to build edges — they only check whether emergent clusters correspond to recognized genres.

Result 1 — genres emerge from prose (robust)

Eight communities emerge, each named by an LLM from its member titles + distinctive vocabulary, and each matched against the held-out labels:

Emergent genre	Distinctive vocabulary	Held-out label agrees	Exemplars
Detective fiction	detective inspector police murder holmes poirot	Detective and mystery stories	Sign of the Four, Leavenworth Case
Science fiction	scientific science machine	Science fiction	Clockwork Man, Tono-Bungay
Nautical adventure	deck cabin mate aboard shore	Adventure	Call of the Wild, Almayer's Folly
Historical romance	thy thou sword knight soldier	Historical fiction	Quo Vadis, She
Early English novel	madam parson behaviour discourse	England — Fiction	Robinson Crusoe, Tom Jones
American realism	car dollars chicago hotel york	Psychological fiction	The Great Gatsby, Age of Innocence

Genres were recovered from the words alone, and independent labels confirm them. See visualize_genres.py → genre_network.html for the interactive network (click any novel or genre to trace its cluster) — live on the site, alongside Phase 2's influence network, under the "Explore the genres" tab. (visualize.py → literary_genres.html still produces the original static two-panel Plotly export.)

Result 2 — there is no global mutation rate, but one genre genuinely emerges

The original thesis wanted a genre-mutation rate over time. There is no such global rate — the apparent signal is a stack of confounds, each of which we found (and were fooled by) in turn (controls.py):

Test	Apparent signal	Verdict
Per-year mutation rate	"accelerates, inflection ~1890"	❌ corpus-density artifact (later eras have more books)
Per-book rate + null model	—	❌ no global signal: real chronology (90 events) ≈ shuffled years (94 ± 15, z = −0.27)
Style-drift control	"science fiction emerges, z = −4.9"	❌ prolific-author artifact (the cluster was H.G. Wells's voice; not robust across k)
+ author control (one book/author) + k/seed sweep	detective fiction, z ≈ −3.0	✅ robust, label-validated

Three structural confounds, each controlled:

1. Corpus density — measure per book, not per year.
2. Style drift — English prose drifts over time (corr(similarity, year-gap) = −0.32), making any text graph temporally structured. Regress the year trend out of the vectors (→ corr ≈ 0).
3. Author voice — ~20–26% of raw k-NN edges are same-author; prolific authors (Wells, Conrad, Doyle: 8–10 books each) form clusters that masquerade as concentrated genres. Use one book per author.

After all three, each emergent community is tested for temporal concentration vs a null (random same-size draws). The result:

• Detective fiction is the one genuine, datable emergence — z ≈ −3.0, concentrated in the books of 1878–1926, matched to the held-out "Detective and mystery stories" label, robust across k = 4–6 and all seeds. This fits literary history exactly: detective fiction is the paradigm genre with a real birth (Poe → Collins → Doyle → the golden age). (An earlier revision of this README said "~1840s–1920s", which read as a corpus figure and was wrong by ~35 years: controls_results.json says 1878–1926. Poe and Collins are literary-historical context, not books in the corpus.)
• Everything else is a perennial mode — gothic, adventure, domestic, historical fiction are spread across all 250 years (z ≈ 0), not "born" at a point.

Conclusion: genre birth is real but rare and genre-specific, not a smooth global rate. The measurable findings are unsupervised genre recovery and the emergence of detective fiction specifically — not a universal mutation rate.

Reproducing

pip install networkx scikit-learn numpy plotly
export GEMINI_API_KEY=...  ANTHROPIC_API_KEY=...   # for corpus assembly only

python build_canon.py        # -> _data/canon.json   (cross-referenced canon)
python build_corpus.py       # -> _data/books.json   (real Gutenberg full text)
python analyze.py            # -> results.json        (communities, naming, null model, sweep)
python visualize.py           # -> literary_genres.html (static two-panel Plotly export)
python visualize_genres.py   # -> genre_network.html (click-to-explore version, needs literary_genres.html + controls_results.json)
EDGE_METHOD=semantic python temporal_network.py   # the year-by-year timeline

Limitations & honest next steps

• Pre-1929 ceiling. Public-domain full text stops at ~1928, so this is the genre story up to modernism — cyberpunk, modern fantasy, and postmodernism are out of reach without a licensed-text or excerpt source.
• The rate question isn't dead, the instrument is. A per-book, null-model-controlled statistic might still find real structure; the raw event count does not. That is the honest open problem.
• Reception vs text. Built — see Phase 3 below. Genre also lives in how readers/critics classify books over time, and a period-reception dataset measures formation directly. This was listed here as the hardest and most direct future dataset; it now exists for the American trade, 1855–1929.

Phase 2 — Author Influence Network

A second, separate result: does textual similarity between specific authors track real, documented influence, beyond chronology alone? Built on the same discipline (measured signal only, held out against real influence claims, never fabricated) but a different graph — directed, author-to-author, chronologically-forward-only candidate edges, two independent similarity signals (stylistic TF-IDF + conceptual embedding) kept separate rather than merged into one score.

build_bibliography.py cross-references 2,411 works across 108 authors (each confirmed by two independent model families); build_corpus.py resolves 583 of those to real Gutenberg prose across 77 authors, Homer through the 1920s; build_influence_graph.py builds 2,915 directed candidate edges, each carrying both scores.

Held out against two independent validation sources — neither used to build a single edge:

validation source	stylistic z	conceptual z
known_influences.json — 130 pairs, LLM-enumerated critical consensus	0.91 (not significant)	9.47 (highly significant)
Wikidata P737 "influenced by" — 102 pairs, no LLM involved	2.45 (significant)	7.16 (replicates)
well-represented subset (n=47 authors, density control)	2.97 (significant)	6.25 (replicates)

Conceptual similarity is a real, replicated result — not a density artifact (the well-represented subset still holds), not a same-form artifact (cross-form documented pairs like Wagner → Nietzsche score the same as same-form pairs). Stylistic similarity is a genuinely open question, deliberately left unresolved: null on the full 130-pair sample but significant in both narrower independent checks. Two honest readings, not adjudicated — either the full-sample null is real and both narrower checks are small-N noise landing the same lucky direction, or the full-sample null was itself partly an artifact that both narrower checks correct. Worth a dedicated pass before claiming or dismissing it either way.

Full design, results, and honest limits in docs/PHASE2_INFLUENCE_NETWORK.md.

Visualization: visualize_influence.py → influence_network.html — an interactive, chronologically-laid-out, directed graph. Click any author and their edges light up; a side panel shows both similarity scores per connection, plus the real citation note where a connection is independently documented (Wikidata or the held-out list), not just measured. Hosted live at aidanjude.vercel.app/research/literature-mutations.html.

Phase 3 — the period reception series

A third result, and the one nobody else has built. Underwood defines genre membership with modern retrospective labels and names that as a limitation. Genre also lives in how the trade classified books at the time — so this measures that directly, from the American book trade's own dated pages.

Source. Two serials on Internet Archive, free OCR, no lending restriction, no OCR cost: American Publishers' Circular and Literary Gazette (1852–1871) spliced ahead of Publishers' Weekly (1872–1929). 3,487 dated issues, 128.5M words, 0 fetch failures. Two probes got here — Gutenberg back matter was a no-go (4 publisher ads in 343 books), and the two trade catalogues the obvious next guess named both failed on measurement.

The artifact is Publishers' Weekly's Weekly Record annotation — one dated descriptive line per American book published. PW states the policy in its own section masthead, which is what licenses this as classification rather than review opinion:

"The annotations are descriptive, not critical; intended to place not to judge the books."

The finding

"Mystery story" appears zero times in 41.8 million words of the American book trade across 1,690 issues, 1855–1895 — then 1,605 times in the following 34 years.

First occurrence 1896, sustained from 1896, take-off 1906–1909, plateau near 40 per million to 1929. It is not one repeated house ad (169 hits across 45 of 56 issues in 1920) and not authors' titles — the hits sit in the trade's own descriptive line: "A mystery story for girls" is a shelf label, not a review.

A second datable genre, found by the reception side alone. And detective fiction now agrees on three independent clocks: the books (1878–1926), the general language (Ngrams take-off 1889), and the trade (sustained 1873, take-off 1891–1897, still rising at 1929).

One honest correction to our own method

A fraction-of-peak take-off computed on a smoothed sparse series can precede the term's first actual occurrence — a 9-year moving average spreads a first spike backwards. It did: sensation novel first appears in 1863 and the estimator reported a take-off of 1859, which happens to match Phase 2's Ngrams date exactly. That would have been a false replication. Across smoothing widths 9/5/3/1, sea story's take-off moves 74 years. Take-offs therefore ship as bands, and a band spanning more than 10 years is reported as not measurable. Five of eleven terms survive the rule.

Limits. It is the American trade, so any regression against the textual series is a US-side measurement; IA's British holding is 24 items. The annotation series only exists from 1878 — the Circular never annotated — which is why the parser-free per-million series is the primary one. 1929 is a hard ceiling on both sides.

Full method, all eleven terms and every caveat: docs/S6-SLICE3-RECEPTION-SERIES.md, with the two probes that found the source in docs/S6-SLICE1-BACKMATTER-PROBE.md and docs/S6-SLICE2-TRADE-CATALOGUE.md. The dated per-book table is _data/reception_entries.jsonl.gz — 1,453 genre-bearing trade annotations, 1878–1929.

python build_reception_series.py     # -> reception_series.json  (69 min, once)
python analyze_reception_series.py   # -> reception_clock_trade.json  (seconds)

References

See docs/PROPOSAL.md for the original 2021 proposal and its full reference list (Stanford Literary Lab's Quantitative Formalism, Moretti, Hope & Witmore's Docuscope work, Galton's Vox Populi, and the clustering literature).