Independent computational analysis

A Stylometric Study of the Book of Mormon

Function-word analysis with the controls the famous studies on both sides skipped: method validation, adjacency blocking, a fake-narrator null, a lesion test, and open-set attribution.

252,099 words of Book of Mormon text (KJV block quotations excluded) · 10 candidate authors, ~460,000 words of candidate/control prose · Burrows’ Delta on 100 most-frequent words · July 2026 · fully reproducible (scripts & corpora in stylometry/)

1 · Does the method work?

Yes. It identifies six known authors of the era from 2,000-word chunks at κ = 0.88 after removing adjacency leakage — and it cleanly tells apart two other deliberate King-James imitations. The instrument is sound.

2 · Do the internal narrators have distinct “wordprints”?

Not detectably. The claimed effect reproduces only under the original studies’ flawed design, which also certifies fake narrators sliced from one narrator’s text (κ 0.61). Properly controlled, narrator identification falls to κ 0.21 vs 0.88 for real authors; Jacob, Moroni, and every embedded speaker drop to ≈0%.

3 · Does it match any 19th-century candidate?

No. 83–86% of chunks are farther from every candidate than genuine same-author text ever is. Spalding’s style is among the farthest of all candidates; Rigdon’s is far too. The forced-choice “winner” is an unrelated Scottish KJV-imitator — register, not authorship.

Read both ways, honestly: stylometry offers no support for multiple ancient authors, and it actively disfavors the Spalding–Rigdon theory. But “matches nobody” cannot rule Joseph Smith in or out either — no candidate left a surviving sample of sustained scripture-pastiche narrative, and a person’s letter voice is not their deliberate-imitation voice. Neither side should cite stylometry as strong evidence.

1Why re-do this at all

The pro–Book of Mormon “wordprint” studies (Larsen, Rencher & Layton 1980; Hilton 1990) compared blocks of the same narrator’s contiguous text against each other, so topic and adjacency could masquerade as authorship. The anti side’s best-known study (Jockers, Witten & Criddle 2008) used closed-set classification that must assign every chapter to someone on its list — a list that omitted Joseph Smith — and concluded Spalding and Rigdon. Holmes (1992) found the internal “authors” cluster together; Schaalje et al. (2011) showed open-set methods match most chapters to nobody. Every one of those disputes gets an explicit control below.

2Validation, then the wordprint claim

Bar chart comparing naive vs blocked attribution kappa for known authors, Book of Mormon narrators, and fake narrators — **The headline result.** Naive testing (light bars) cannot tell real voices from fake ones: Book of Mormon narrators score 0.72, but arbitrary contiguous slices of a *single* narrator’s text score 0.61. Blocked testing (dark bars) separates them: real authors keep κ 0.88, Book of Mormon narrators fall to 0.21, fake narrators to chance. The 1980-style wordprint signal is mostly adjacency leakage.

Test (2,000-word chunks, 100 MFW)	Naive κ	Blocked κ
Six known real authors (incl. two KJV imitators)	0.95	0.88
Book of Mormon narrators (Nephi, Jacob, Mormon, Moroni)	0.72	0.21
Fake narrators sliced from Mormon’s own text (null)	0.61	−0.18

Blocked per-narrator recall: Mormon 0.76 (the dominant class), Nephi 0.37, Jacob 0.00, Moroni 0.00 — and in the speaker-level test replicating the 1980 design, Benjamin, Abinadi, Alma, Amulek, and Samuel all score 0.00. The weak residual (κ≈0.2, mostly Nephi-vs-Mormon) is as consistent with topic difference between the small plates and the abridgment, or drift over a year of dictation, as with different authors.

Two PCA scatter plots: six known authors form separated clusters; Book of Mormon narrators are intermixed — **What separation looks like.** First two principal components (~20% of variance; the full analysis uses all 100 dimensions). Left: known authors form real clusters — even Hunt and Linning, both writing KJV pastiche. Right: the Book of Mormon’s narrators, same method — largely one intermixed cloud.

3The 19th-century candidates

Ten candidates: Joseph Smith (holograph/dictated documents only, from Joseph Smith Papers verbatim transcripts), Sidney Rigdon (74k words of signed prose), Oliver Cowdery, Solomon Spalding (the actual Oberlin manuscript), Parley Pratt, W. W. Phelps, David Whitmer, Ethan Smith (View of the Hebrews), Gilbert Hunt (The Late War), and Michael Linning (First Book of Napoleon, 1809 — a KJV-imitating text by a Scottish lawyer with no connection to Mormonism, included as a register control). Distractors: Irving, Cooper, and the KJV itself.

Dot plot of Burrows Delta distance from Book of Mormon chunks to each candidate; all candidates sit in the different-author range; Spalding near the bottom — **Nobody is close.** Every candidate’s distance from the Book of Mormon sits in the calibrated different-author range (grey), far outside the same-author range (green). The nearest “candidate” is the register control — and **Spalding is farther from the Book of Mormon than almost anyone tested**, with Rigdon in the bottom half. Robust to removing archaic marker words (unto, yea, behold…) from the features.

Histogram showing Book of Mormon chunks' nearest-candidate distances fall in the different-author distribution — **The open-set test.** Green: how far chunks of known authors sit from their own author’s profile. Violet: each Book of Mormon chunk’s distance to its *nearest* candidate. 83–86% of chunks lie beyond the same-author 95th percentile — the honest answer is “matches nobody on the list,” not a name.

Why the 2008 Spalding–Rigdon result happened — the lesion test. Closed-set classifiers must pick someone. Deleting Cooper from our candidate list and re-attributing his chunks sends 62 of 74 confidently to Irving. Forced to choose for the Book of Mormon, the votes scatter — the “winner” is Linning (36 of 123), the register control, with Spalding receiving zero. A method that crowns an unrelated Scottish lawyer is measuring biblical register, not authorship.

What the style actually is: massively over-used came (“it came to pass,” z = +2.9), should, that, they, unto, people, now, many, them, land; under-used articles and prepositions (a, an, in, on, at, from, with). A single sustained oral-formulaic KJV pastiche, unlike any candidate’s natural prose — including, for what little his 20k-word corpus can show, Joseph Smith’s own letters.

4Limitations — read before quoting

The register confound cuts both ways. Nobody in the pool left a sustained scripture-pastiche narrative in his own hand; a letter-writing voice is not an imitation voice. “Matches nobody” does not prove no 19th-century person wrote it.
Joseph Smith’s reliable corpus is small (~20,000 words, holograph/dictated only) and genre-mismatched. His distance should not be over-read in either direction.
Transmission noise: scribes, compositors, OCR in four candidate corpora, the Pratt text following an 1881 edition, Whitmer writing in 1887. Function words are robust to most of this, but it is not zero.
Narrator maps are chapter-level approximations; both narrator- and speaker-level schemes were tested and agree.
The weak residual narrator signal (κ≈0.2) is real but unattributable — authorship, topic, genre, and dictation drift are confounded within a single translated/dictated text.

5Bottom line for your document

The stylometric argument for the Book of Mormon (distinct wordprints) fails its own controls: the design that produces it also finds “distinct authors” inside text known to have one narrator. The best-known stylometric argument against it (Spalding–Rigdon attribution) also fails: by direct distance measurement its two heroes are among the worst matches available, and the closed-set method that produced it misattributes 84% of a known author’s text when the true author is off the list. What remains, stated plainly: the Book of Mormon is stylometrically one sustained pseudo-biblical idiom that cannot be pinned on any tested author — a result that supports neither the ancient-authors claim nor a specific modern-author claim, and leaves the book’s origin to be argued on other evidence.

Methods: MFW relative frequencies (50/100/200 tested), Burrows’ Delta, leave-one-out nearest-neighbour with same-book/segment blocking, Cohen’s κ, centroid attribution, open-set calibration at the same-author 95th percentile; sensitivity runs without archaic markers. Corpora: Book of Mormon (bcbooks/scriptures-json; 14,839 words of KJV block quotation excluded), Joseph Smith Papers verbatim transcripts, sidneyrigdon.com and centerplace.org human transcriptions, Wikisource, Project Gutenberg, archive.org. Prior literature: Larsen, Rencher & Layton, BYU Studies 20 (1980); Hilton, BYU Studies 30 (1990); Holmes, JRSS-A 155 (1992); Jockers, Witten & Criddle, LLC 23 (2008); Schaalje, Fields, Roper & Snow, LLC 26 (2011). Full write-up: stylometry/REPORT.md; per-document provenance in corpus/raw/*/MANIFEST.md.