# The "Beale" letters and the 1885 narrative: computational stylometry and language dating

All code is in `code/` (prefixes `stylo_`, `lang_`), all machine-readable outputs in `results/`, the
control corpus (raw downloads, cut samples, manifest with source URLs) in `data/controls/`.
Every number below can be regenerated by running the scripts listed in "Reproducibility".

## 1. Conclusions first

**A. Same hand?** Yes, with high probability. On every distance measure that discriminates authors
reasonably well in this corpus (Burrows/Cosine Delta on 200-500 most frequent words, pronoun-free
Delta, character 4-grams), the three "Beale" letters (2,716 words, dated 1822) and the anonymous
author's own narrative (3,954 words, written c. 1885) are closer to each other than **99-100 % of the
6,352 pairs of samples by different control authors**, and as close as typical pairs of samples by the
**same** author. With the texts cut to the same length as the controls (2,000 words each), the
letters-narrative distance is matched by 0.2-1.0 % of different-author pairs and by 15-31 % of
same-author pairs; the likelihood ratio (same hand : different hands) from kernel-density estimates is
about 10-25 for the length-matched pair and 70-650 for the full-length pair. Pronoun-free and
function-word-only variants give the same answer, so the result is not driven by shared subject
matter or by the first-person register of letters. Nine pairs of different control authors who wrote
on the *same topic* (e.g. Gregg vs. Thomas James on the Santa Fe trade, three Lynchburg local
historians, Cooke's Life of Lee vs. Lee's own letters) all come out farther apart than the Beale pair.
The narrative is the nearest "author" to the letters by a wide margin (Cosine Delta 0.69 vs. 0.90 for
the runner-up, Jefferson), and the letters are the nearest text to the narrative (0.55 vs. 0.85 for
the runner-up, Eggleston).

Turning the likelihood ratios into a probability requires a prior. With even odds a priori, the
length-matched results give P(same hand) ≈ 0.9-0.96; the full-length results give > 0.98. I would
state the result as: **the observed stylistic closeness would arise between two different authors in
roughly 1 case in 100 to 1 in 500 in a corpus of 27 nineteenth-century American authors, but is
ordinary for two texts by one author.** The important non-stylometric caveat is that the anonymous
author admits he "read over and over again the letters ... endeavoring to impress each syllable they
contained on my memory", so an alternative reading (a hoaxer imitating letters he had memorised, or
a narrator whose prose was contaminated by them) cannot be excluded by stylometry alone — but the
shared traits are habits nobody imitates deliberately (rates of *until*, *myself*, *such*, *may*,
*each*, untriggered reflexive pronouns, negative infinitival passives), which is what one expects
from a single writer.

Specific corroboration of Nickell/Pival: the "misuse of reflexive pronouns" that Jean Pival reported
in 1982 is real and rare. Untriggered reflexives (*like myself*, *not so reliable as yourself*, *for
myself and family*, *as sanguine as himself*, *more fortunate than myself*) occur 3.7 and 2.3 times
per 1,000 words in the letters and the narrative; the control mean is 0.4 (sd 0.5) and **only 1 of 115
control samples (a Jefferson letter sample) reaches the lower of the two Beale rates.** The
construction *not/never to be + participle* (*not to be delivered*, *never to be realized*, *not to be
resisted*, *not to be numbered with the living*) is at 0.4-0.5 per 1,000 in both against a control
mean of 0.05. On the other hand the crude punctuation/syntax rates that summaries of Nickell's article
emphasise (semicolons, relative clauses, infinitives, conjunctions, sentence length) barely
discriminate authors at all in this calibration (likelihood ratio ≈ 0.7-0.9, i.e. uninformative), and
the two texts actually *differ* on semicolons (1.5 vs 7.1 per 1,000) and on *that* (8.5 vs 15.2).

**Poe** is excluded. The Gold-Bug's nearest neighbours are Bagby, Cable and Twain; Poe's letters'
nearest neighbours are Twain's letters, Lee and Randolph; Poe's samples sit at ordinary
different-author distances from both Beale texts (e.g. 0.95 and 0.90 on the pronoun-free measure,
against 0.55-0.70 for letters vs. narrative). This agrees with Poundstone's 1983 test as summarised on
Wikipedia.

**Morriss's quoted statement** (865 words) is as close to the narrative as the two halves of the
narrative are to each other (0.61 vs 0.65, pronoun-free Cosine Delta; both at the 0.1 percentile of
different-author pairs). This is expected — the author says he "reduced [it] to writing" — and it
serves as an internal control: a passage the narrator admits to having written in his own words
behaves exactly like the "Beale" letters do.

**B. Language of 1822 or later?** The letters contain a small but, by calibration, abnormal amount of
vocabulary that was rare or unknown in American print before 1822 and common after 1850:

| item in the letters | evidence | verdict |
|---|---|---|
| *stampeding* ("securing many and stampeding the rest") | Earliest checkable English use I could verify: a deposition in the Austin Papers (Texas, written c. Oct. 1823, about events of Aug. 1822): "the Cavellada Stampeded ... the stampede etc. ... the Cavellada that stampeded" (Annual Report of the AHA for 1919, vol. II pt 1, p. 699, publ. 1924; this is the OED's 1823 verb citation, cf. Etymonline "verb 1823"). Noun forms: OED earliest 1828 ("stampiado"), Merriam-Webster noun 1828, verb 1838; Timothy Flint's novel *Francis Berrian* (1826) has "took what the Spanish call the 'stompado'" — verified in reprints of the passage in the *Oxford Observer* (Paris, Me.) 2 Nov. 1826, *Catawba Journal* 7 Nov. 1826, *Saturday Evening Post* 25 Nov. 1826, *The Casket* Jan. 1827, and in school readers from 1826 on, always with the word marked as Spanish; "A Stampedo" is the title of a soldier's narrative in the *Military and Naval Magazine of the United States*, Dec. 1834 (vol. 4 no. 4, pp. 296-98; contents advertised in the *Daily National Intelligencer* 22 Dec. 1834), which ends by explaining that the word "is a North American attempt to adopt a Spanish word into our language ... in general use among those, who have witnessed similar occurrences" — i.e. still a frontier term needing a gloss in 1834; Gregg 1844 prints "*stampeded*" in italics as a frontier term; Thomas James 1846 (about 1821-22) has "stampedo/stampede". American newspapers (Chronicling America, every hit OCR-verified, section 3.3): no *stampede/stampeded/stampeding* in 1800-1835; first *stampede* 12 Aug. 1836 (*Arkansas Advocate*), then 1839 (Comanches stealing horses "by stampede"); first *stampeded* and *stampeding* July-Aug. 1842 (*Southern Argus*, letters from the Texan Santa Fe expedition); figurative uses from 1843-44. Google Books: essentially zero before 1840, rising after 1845; the Wikipedia claim of an 1834 "stampeding" cites M. E. W. Sherwood's *Manners and Social Usages*, a book of 1884-87, misdated. | **Not a hard anachronism, but at the extreme early edge of the word's history.** A man who had spent 1817-21 around Santa Fe could have carried the Spanish *estampida* home in Jan. 1822, 20 months before the earliest surviving English use; but the word reached print as *stompado/stampedo* (1826-34) before *stampede* (1836), and the fully anglicised participle *stampeding* is unattested until 1842. Nickell's dates (1844; "stampeding unattested before 1883") are too late. |
| *improvised* ("such tools and appliances as they had improvised") | Merriam-Webster first use 1788, Etymonline 1808 (sense: compose or perform extempore, from *improvvisatore*); Google Books frequency before 1822 is one-seventh of the 1850-1900 level. Chronicling America 1800-1835 (25 + 25 hits OCR-checked): the verb first appears in 1830 and only in the extempore-composition sense ("an improvised account of given chapters in philosophy" 1830; Scott "literally improvised" 1833; "improvised this piece of pleasantry" 1835); the earlier "hits" are OCR errors for *improving*. A further window for 1836-1845 (8 verified uses, 1837-40) is the same: improvised prayers, verses, songs. The concrete sense "contrive from what is at hand" is the later one; I could not date it independently of the OED (paywalled). | **Suggestive, unproven.** The word existed (rarely, for extempore verse and music); the letters' sense is unattested in this newspaper corpus through 1840. |
| *grizzlies* ("encounter the savage grizzlies") | "Grizzly bear" is recorded from 1806 (Lewis and Clark; Etymonline). The clipped noun *grizzly/grizzlies* is "early 19th century" per Merriam-Webster with no date; Google Books has *grizzlies* at noise level until the 1850s (ratio pre-1822/1850-1900 = 0.026); Chronicling America: none of the 40 search hits for 1800-1835, nor of the 40 earliest hits for 1836-Feb. 1838, is *grizzlies* (all *grizzly*, *grizzled*). | **Probable anachronism** for a Virginian of 1822; the plural clipping is a mid-century Western colloquialism. |
| *objective point* ("our objective point being Santa Fé") | Military term from French *point objectif*; Etymonline 1852. Chronicling America has no verified use in 1800-1846; the earliest verified printed uses are Mexican-War strategy commentary of January 1847 — *Richmond Daily Whig*, 5 Jan. 1847, p. 2 ("the objective point, San Luis, might have been gained"), then the *Daily Union* (Washington) 9 Jan., *Staunton Spectator* 14 Jan., *Alexandria Gazette* 15 Jan. and *Boon's Lick Times* 23 Jan. 1847 ("San Luis should then have been, in the plan of campaign, the objective point") — six newspapers in three weeks and none before (results/lang_attest_window_objective_point_1836-1848.csv). Google Books ratio 0.029, take-off 1861-65 (Civil War). | **New, strong anachronism candidate** (not in Nickell). A hunting party's "objective point" is Mexican-War/Civil-War officer's English, which reached Virginia newspapers in 1847. |
| *reliable* (twice) | Etymonline: "not common before 1850, and often execrated thereafter in Britain as an Americanism"; Google Books ratio 0.076, take-off 1845-55. Chronicling America 1800-1835: the four pre-1812 hits are OCR-doubtful (e.g. 1804 "reliable ... to a suit" = *liable*); no unambiguous use verified before 1822. | **Weak-to-moderate.** |
| *appliances* (tools and appliances) | General sense old (MW 1555); the "apparatus" sense and the plural are 19th-century; ratio 0.071. Chronicling America 1815-1822: three uses quoting Shakespeare's "appliances and means to boot" and one general use (1822). | **Weak.** |
| *locality*, *everything* (one word), *in the meantime* (one word), *pretense*, *-or* spellings (*honor, favorably, laborers*) | Later-favoured forms: before 1822 American books print *every thing* 6:1, *in the mean time* 8:1, *pretence* 21:1, and the *-our* spellings still slightly outnumber *-or* in the early control samples. The letters use the later form in every case. | **Weak individually** — an 1885 compositor could have modernised spelling and compounding — but see the calibration below. |

Calibrated against twelve 2,716-word samples of genuine American writing of 1805-1843 (Jefferson,
Randolph, Jackson and J. Q. Adams from the 1820s; Edwin James, Flint, Hall and Paulding from
1817-28; Pike, Biddle, Irving and Gregg), the letters have **4-5 times the normal share of "late"
vocabulary**: 0.85 % of their content-word types have a pre-1822 Google-Books frequency below one
tenth of their 1850-1900 frequency, against a control mean of 0.19 % (sd 0.16, maximum 0.47 %; z =
+4.1; 0 of 12 controls as high). Two words (*stampeding*, *grizzlies*) are "strong" (later common,
essentially unattested before 1822); the only control with two such words is Gregg's 1844 Santa Fe
narrative (*ranchero, vaqueros*). On a looser threshold (ratio < 0.25) Gregg 1844 and Irving 1835 come
level with the letters, so the honest reading is that the letters' vocabulary resembles a Western
narrative of the late 1830s-1840s or after, and is unlike every 1810-1828 text in the set (0.0-0.5 %
on the strict measure). The count is small in absolute terms (a handful of words), which is why the
individual items are dated below.

The reverse check (deliberate archaising) finds a mixture: early-favoured items are present —
*&c.* (twice), *enquiring*, *such an one*, *winter quarters*, *punctillio* and *exhilerating*
(spellings already old-fashioned in 1822), *the States*, the capitalised season *Fall* — alongside the
later-favoured items above. All of these, and the idioms *well nigh*, *worth the candle*, *ho for the*,
*encounter the expense* and *personating*, are verified in American newspapers of 1800-1835 (section
3.3), many of them in Richmond papers. That pattern fits an 1880s writer lightly archaising (or an 1822 manuscript
modernised by an 1885 printer) better than it fits an unretouched 1822 letter; the lexical items that a
compositor could not have introduced (*objective point*, *grizzlies*, *stampeding*) are the decisive
ones. Function-word style, interestingly, cannot date the letters: on nearest-period classification
(72-80 % accurate on controls) the letters look *early* and the narrative *late*, and within the
letter genre the letters are nearer to Jefferson/Randolph/Jackson than to Lee/Twain/Poe. The
epistolary register of 1822 was imitable; the vocabulary of 1822 was not.

**Overall.** The two lines of evidence are independent and point the same way: the letters were most
probably composed by the pamphlet's author, in the 1880s, in a deliberately older epistolary register.
Confidence: high for common authorship (≈ 95 % on the calibrated stylometry alone, before considering
the anachronisms), moderate-to-high for a post-1822 date of composition (the calibrated anachronism
density is > 4 sd above genuine texts, but the strongest single items are "probable", not "certain",
anachronisms). The OCR-verified newspaper record sharpens the picture without closing it: in 1834 a
military writer still had to explain *stampedo* as "a North American attempt to adopt a Spanish
word", *stampede* reaches American newspapers in 1836 and *stampeding* in 1842, *grizzlies* is
unattested before 1838, *objective point* before 1847, and *improvised* means only "composed
extempore" through 1840 — while every archaism in the letters is ordinary newspaper English of
1800-1835. A Virginian writing in January 1822 would have had to be ahead of the printed record on
three or four independent words at once.

## 2. What is new relative to Nickell (1982)

Nickell, "Discovered: The Secret of Beale's Treasure", *Virginia Magazine of History and Biography*
90:3 (July 1982) 310-324 (JSTOR 4248566, paywalled; I did not see the article and rely on secondary
summaries — see Sources) argued (i) impressionistically, with a tabulation by the linguist Jean G.
Pival of syntactic features (sentence length; shares of *the*, *and*, *of*; reflexive-pronoun
"misuse"; negative passives such as "never to be told", "never to be realized"), that Beale and Ward
are one writer, and (ii) that *stampede/stampeding* and *improvise(d)* were unknown before the 1840s.
This note adds:

1. A calibrated, reproducible authorship test: 27 control authors, 115 samples, four families of
   features, empirical same-author and different-author distance distributions, leave-one-out
   attribution accuracy, author-level bootstrap, pronoun-free and function-word-only controls for
   genre, and same-topic different-author controls for subject matter. The verdict is the same as
   Nickell's but now carries a number (≈ 1 % of different-author pairs are as close) instead of an
   impression.
2. Quantification of Pival's reflexive-pronoun feature (1 of 115 control samples matches) and a
   negative result for the punctuation/syntax rates (uninformative).
3. Correction of the anachronism dates: *stampede(d)* is in English by 1823 (Austin Papers, Texas)
   and in print by 1826-28; *improvise* by 1788/1808. Neither is a hard anachronism for 1822. An
   OCR-verified newspaper chronology (Chronicling America, 597 + 183 hits checked) shows, however, that
   *stampede* entered American newspapers only in 1836 and *stampeding* in 1842, that *improvise(d)*
   before 1836 meant only "compose extempore", and that every "archaic" item in the letters was
   ordinary newspaper English of 1800-1835.
4. Two new candidate anachronisms — *objective point* (earliest verified use Richmond, 5 Jan. 1847;
   Civil-War vogue) and
   *grizzlies* (1850s) — plus *reliable* (post-1850 vogue), and the compounding/spelling pattern.
5. A calibration of the anachronism density against genuine 1805-1843 texts (the letters are 4-5×
   normal, z ≈ +4).
6. A negative result on style-based dating: by function words the letters look like 1820s letters,
   i.e. the writer's register imitation was good; vocabulary betrays him, style does not.
7. Poe tested and excluded quantitatively; Morriss's statement shown to behave as the narrator's prose.

## 3. Tables

### 3.1 Authorship: distances and calibration

Control corpus (data/controls/manifest.csv): 115 samples of ~2,000 words from 27 authors —
American 1805-1846: Jefferson (letters 1816-26), John Randolph (letters 1806-32), Andrew Jackson
(letters 1820-28), J. Q. Adams (diary 1819-21), Edwin James (1823), Timothy Flint (1826), James Hall
(1828), J. K. Paulding (1817), Pike (1810), Biddle's Lewis & Clark (1814), Irving (1835), Gregg
(1844), Pattie (1831), Thomas James (1846); Virginian/Southern 1858-1904: Bagby, J. E. Cooke (history
1883 and biography 1871), Jefferson Davis, T. N. Page, G. W. Cable, G. C. Eggleston, Mark Twain (narrative
1883 and letters 1876-85), R. E. Lee (letters 1861-70), M. A. Cabell (Lynchburg 1858), W. A. Christian
(Lynchburg 1900), E. Pollock (Lynchburg 1887), Southern Historical Society Papers (1884); Poe (The
Gold-Bug 1843, letters 1829-49). Z-scores are computed on the control samples only.

Distance of the Beale letters to the narrative, and where it falls among control pairs
(full table in results/stylo_delta_report.md; "pct diff" = share of different-author control pairs at
least as close; "pct same" = share of same-author pairs at least as close; LR = kernel-density
likelihood ratio same:different):

| method | LOO accuracy | letters (2716 w) vs narrative (3954 w): pct diff / pct same / LR | both cut to 2000 w: pct diff / pct same / LR | narrative half A vs B (same author): pct diff | Morriss vs narrative: pct diff |
|---|---|---|---|---|---|
| Cosine Delta, 100 MFW | 0.58 | 2.0 % / 35 % / 6.5 | 5.3 % / 51 % / 3.6 | 1.2 % | 0.7 % |
| Cosine Delta, 200 MFW | 0.63 | 0.3 % / 15 % / 25 | 1.0 % / 31 % / 13 | 0.3 % | 0.1 % |
| Cosine Delta, 300 MFW | 0.63 | 0.0 % / 10 % / 73 | 1.0 % / 31 % / 11 | 0.3 % | 0.0 % |
| Cosine Delta, 500 MFW | 0.68 | 0.0 % / 8 % / 305 | 1.9 % / 37 % / 8.5 | 0.1 % | 0.3 % |
| Cosine Delta, 300 MFW, pronouns removed | 0.60 | 0.0 % / 5 % / 645 | 0.5 % / 22 % / 25 | 0.1 % | 0.1 % |
| Cosine Delta, 150 function words only | 0.61 | 0.4 % / 17 % / 22 | 2.1 % / 38 % / 6.6 | 0.8 % | 0.1 % |
| Burrows Delta, 300 MFW | 0.69 | 0.0 % / 4 % / 284 | 4.9 % / 43 % / 2.7 | 0.7 % | 18 % |
| Cosine Delta, 2000 char 4-grams | 0.65 | 0.0 % / 8 % / 147 | 0.2 % / 18 % / 25 | 0.1 % | 0.1 % |
| raw cosine, char 4-grams (no z-scoring) | 0.53 | 4.6 % / 33 % / 1.9 | 25 % / 54 % / 1.0 | 12 % | 21 % |

LOO accuracy = share of control samples whose nearest neighbour is another sample of the same author
(27 authors, 2,000-word samples; chance ≈ 0.04). Same-author pairs n = 203, different-author pairs
n = 6,352. The raw (un-standardised) char-gram cosine and the 100-MFW Burrows Delta are the weakest
discriminators and give the weakest result; every measure with LOO accuracy ≥ 0.6 puts the
length-matched Beale pair at or below the 2nd percentile of different-author pairs.

Author-level bootstrap (300 resamples of the 27 control authors; 95 % interval for the share of
different-author pairs at least as close as the length-matched Beale pair): Cosine Delta 200 MFW
0.4-1.9 %; 300 MFW 0.4-1.8 %; 500 MFW 1.1-3.0 %; 300 MFW without pronouns 0.2-1.0 %; 150 function
words 1.1-3.4 %; char 4-grams 0.0-0.4 %; Burrows Delta 300 MFW 2.3-7.3 %; the two weak measures
(100-150-MFW Burrows Delta, raw char-gram cosine) 8-34 %.

Nearest control authors (mean distance to the author's samples; pronoun-free Cosine Delta, 300 MFW):
letters (2000 w) → narrative 0.689, Jefferson 0.901, Gregg 0.904, Eggleston 0.953, Jackson 0.957;
narrative → letters 0.552, Eggleston 0.853, Cabell 0.890, Jackson 0.893, Lee 0.899;
Morriss → narrative 0.679, Randolph 0.910, Jackson 0.913.

Same-topic, different-author control pairs (pronoun-free Cosine Delta; mean over the 16 sample pairs,
minimum in brackets): Gregg vs Thomas James 0.990 (0.813); Cabell vs Christian 0.848 (0.760);
Cabell vs Pollock 0.924 (0.846); Christian vs Pollock 0.892 (0.628); Cooke's Life of Lee vs Lee's
letters 0.959 (0.819); Davis vs Cooke 0.914 (0.745); Eggleston vs Bagby 0.959 (0.739); Edwin James vs
Pike 0.912 (0.717); Flint vs Hall 0.956 (0.859). Beale letters vs narrative (2000 w each): 0.699.

Function words that make the letters and the narrative alike (z-scores against the controls, letters /
narrative): *until* +2.7/+3.6, *myself* +1.5/+3.2, *such* +2.7/+1.5, *may* +1.7/+2.0, *each*
+3.9/+0.8, *will* +3.4/+0.6, *be* +3.5/+0.5, *might* +1.9/+1.0, *this* +1.4/+1.3, *once* +2.6/+0.6,
*can* +0.9/+1.6; both texts are low on *on*, *about*, *very*, *both*, *yet*, *down*. They differ, as
a letter and a narrative must, on *you/our* (letters high) and *he/his/him/could/nor* (narrative high).

### 3.2 Punctuation and syntax (results/stylo_features_report.md)

Rates per 1,000 words; control mean (sd) over 115 samples; "reach" = number of control samples that
reach the lower of the two Beale values.

| feature | letters | narrative | Morriss | controls | reach |
|---|---|---|---|---|---|
| semicolons | 1.5 | 7.1 | 7.0 | 7.9 (7.4) | — |
| colons | 1.1 | 2.0 | 0.0 | 1.6 (1.8) | — |
| dashes | 2.2 | 0.8 | 0.0 | 4.0 (5.5) | — |
| parentheses / questions / exclamations | 0 / 0 / 0 | 0 / 0 / 0 | 0 / 0 / 0 | 1.0 / 1.4 / 1.2 | 50 / 54 / — |
| *&c.* | 0.7 | 0.0 | 0.0 | 0.1 (0.5) | — |
| relativisers *which/who/whom/whose* | 11.4 | 13.4 | 10.5 | 9.1 (4.0) | 30 |
| *which* | 8.1 | 8.6 | 5.8 | 5.6 (3.0) | 22 |
| *that* | 8.5 | 15.2 | 10.5 | 11.0 (4.9) | — |
| infinitives (*to* + verb, heuristic) | 23.2 | 19.5 | 10.5 | 17.0 (5.6) | — |
| *and/but/or/nor* | 48.2 | 47.9 | 66.3 | 44.3 (7.8) | — |
| *however* | 1.1 | 2.0 | 1.2 | 0.6 (0.6) | 17 |
| *doubtless* | 1.1 | 0.3 | 0.0 | 0.1 (0.2) | 12 |
| *until* | 2.2 | 2.8 | 7.0 | 0.5 (0.6) | 5 |
| passives (*be* + participle) | 18.0 | 13.9 | 14.0 | 10.0 (4.1) | 15 |
| negative passives | 1.1 | 2.3 | 1.2 | 0.7 (0.7) | 22 |
| *not/never to be* + participle | 0.4 | 0.5 | 0.0 | 0.05 (0.1) | 7 |
| reflexive pronouns | 4.8 | 3.0 | 3.5 | 1.8 (1.2) | 13 |
| untriggered reflexives (after *than, like, as, for, to, of* …) | 3.7 | 2.3 | 3.5 | 0.4 (0.5) | **1** |
| *the* / *and* / *of* (% of words) | 5.4 / 3.9 / 3.5 | 6.1 / 3.6 / 3.7 | 4.5 / 5.1 / 2.6 | 7.2 / 3.5 / 4.2 | — |
| sentence length mean (sd) | 30.2 (14.7) | 35.2 (18.5) | 27.7 (12.9) | 27.2 (7.2) | — |
| sentences ≥ 40 words (%) | 23 | 29 | 13 | 20 (11) | — |
| paragraph length (words) | 101 | 107 | 124 | 109 (74) | — |

Standardised distance over all these features: the Beale pair falls at the 35th percentile of
different-author pairs and the 73rd of same-author pairs (LR 0.74): these coarse rates do not
discriminate authors (same-author mean 1.09, different-author 1.36, sd ≈ 0.45). The Pival-type
features are the exception (untriggered reflexives, *not/never to be* + participle, *until*).

### 3.3 Language dating

Words of the letters + cipher-2 plaintext (875 types) with Google Books (American English, 2019
corpus, no smoothing) mean frequency 1780-1821 below one quarter of the 1850-1900 mean
(results/lang_ngram_letters.csv; proper nouns excluded):

| word | ratio pre-1822 : 1850-1900 | years with any use 1780-1821 | word | ratio | years |
|---|---|---|---|---|---|
| stampeding | 0.017 | 2 (noise) | tethered | 0.20 | 24 |
| grizzlies | 0.026 | 3 (noise) | ravine | 0.21 | 33 |
| *objective point* (phrase) | 0.029 | 9 (noise) | showing | 0.21 | 42 |
| appliances | 0.071 | 29 | excitement | 0.22 | 40 |
| reliable | 0.076 | 32 | accordance | 0.22 | 40 |
| pretense | 0.082 | 27 | location | 0.22 | 41 |
| locality | 0.095 | 37 | excavation (cipher 2) | 0.24 | 30 |
| grizzly | 0.13 | 24 | imperative | 0.24 | 33 |
| residences | 0.13 | 33 | shipment | 0.25 | 38 |
| improvised | 0.14 | 19 | | | |
| favorably | 0.15 | 40 | | | |
| buffalo | 0.15 | 39 | | | |
| laborers | 0.17 | 38 | | | |
| bracing | 0.19 | 32 | | | |
| everything | 0.20 | 42 | | | |

Phrases and spellings checked separately (results/lang_ngram_phrases.csv): *ho for the* 0.65,
*worth the candle* 0.26 (the proverb is 17th-century), *well nigh* 0.78, *personating* 1.2,
*punctilio* 2.3 / *punctillio* 2.0, *exhilerating* 2.5, *enquiring* 8.3, *such an one* 2.1, *winter
quarters* 4.0, *every thing* 3.8 vs *everything* 0.20, *in the mean time* 3.1 vs *in the meantime*
0.51, *pretence* 3.7 vs *pretense* 0.08, *favourably* 3.6 vs *favorably* 0.15, *labourers* 4.2 vs
*laborers* 0.17, *Santa Fe* 0.26 vs *Santa Fé* 0.10 (the accented form is a mid-century printing
habit, not an archaism), *buffaloes* 0.70, *outfits* 0.28.

Calibration (results/lang_ngram_calibration.md), content-word types only:

| text (2,716 words) | types | % with ratio < 0.1 | % with ratio < 0.25 | % "strong" (ratio < 0.05 and later common) |
|---|---|---|---|---|
| Beale letters + cipher 2 | 708 | **0.85** | **3.39** | **0.28** (stampeding, grizzlies) |
| 12 genuine 1805-1843 samples, mean (sd) | 551-866 | 0.19 (0.16) | 1.62 (0.96) | 0.03 (0.07) |
| worst control | | 0.47 | 3.46 | 0.22 (Jefferson: *egoism*) |
| z of the letters; controls ≥ letters | | +4.1; 0/12 | +1.8; 1/12 | +3.6; 0/12 |
| anonymous narrative, 1885 (3,954 w, for comparison) | 960 | 0.2 | 1.5 | 0.00 |

The measure is topic-sensitive: the letters' "late" words are mostly frontier words (*stampeding,
grizzlies, buffalo, ravine, tethered, outfits, appliances*) whose printed frequency exploded with
westward expansion after 1840, and the 1885 narrative, whose subject is not the frontier, scores
*lower* than the letters. The relevant comparison is therefore with the seven frontier/Western
controls in the calibration set, all written by men who had been on the plains or the frontier:
Biddle 1814 (0.1 % / 1.2 %), Pike 1810 (0.2 / 1.3), Edwin James 1823 (0.5 / 2.8), Flint 1826 (0.1 /
1.0), Hall 1828 (0.0 / 2.0), Irving 1835 (0.4 / 3.2) and Gregg 1844 (0.3 / 3.5; "strong" words
*ranchero, vaqueros*). On the strict measure (ratio < 0.1) the letters (0.85 %) exceed every one of
them; on the loose measure (ratio < 0.25) Gregg's 1844 Santa Fe narrative matches them and Irving's
1835 tour comes close. The fair summary is that the letters' vocabulary profile resembles a Western
narrative of the 1835-1845 vintage (or later) more than any text of 1810-1828 in the set — the
strict-measure gap between the letters and the genuine 1820s texts (Jefferson, Randolph, Jackson,
Adams, James, Flint, Hall, Paulding: 0.0-0.5 %) is the load-bearing comparison, and the z-scores
above should be read with the small n (12) and the non-normal spread in mind.

Orthographic conventions (results/lang_spelling_conventions.md; early : late form):

| convention | letters | narrative | early controls 1805-46 (116,700 w) | late controls 1858-1904 (106,900 w) |
|---|---|---|---|---|
| every thing : everything | 0 : 2 | 0 : 0 | 14 : 4 | 1 : 24 |
| any thing : anything | 0 : 0 | 0 : 0 | 17 : 7 | 1 : 26 |
| mean time : meantime | 0 : 1 | 0 : 0 | 6 : 2 | 0 : 9 |
| &c. : etc. | 2 : 0 | 0 : 0 | 29 : 14 | 5 : 9 |
| -our : -or (honour/honor, favour, labour…) | 0 : 8 | 0 : 5 | 87 : 72 | 11 : 137 |
| enquir- : inquir- | 1 : 0 | 0 : 0 | 17 : 15 | 2 : 11 |
| -ence : -ense (pretence, defence, offence) | 0 : 1 | 0 : 0 | 21 : 1 | 9 : 9 |

**Verified early attestations.** `code/lang_attest_search.py` ran a phrase search for 22 candidate
words and phrases in Chronicling America (loc.gov JSON API, issues dated 1800-1835) and checked every
one of the 597 search hits against the page OCR (results/lang_attestations.csv, .md). The search
engine matches fuzzily — "stampede" returns "stamped, or painted", "grizzlies" returns "grizzly" and
"grizzled", "improvise" returns "improving the complexion" — so only OCR-verified hits count, and a
word broken across an OCR line can be missed (a false negative, never a false positive).

| item | verified in 1800-1835 newspapers? | earliest verified uses (page OCR checked; URLs in Sources) |
|---|---|---|
| stampede, stampeded, stampeding | **no** (2 fuzzy hits each, all "stamped") | Earliest verified newspaper uses at all (date-ordered window search 1826-1845, results/lang_attest_window_stampede_1826-1845.csv, 22 verified of 40): *Arkansas Advocate* (Little Rock) 12 Aug. 1836, figurative, "you took a stampede, and rushed away from us"; *New-Hampshire Statesman* 23 Mar. 1839, "a few Comanches stole silently into the droves of horses while feeding at night, and by stampede recaptured the whole"; *Southern Argus* (Columbus, Miss.) July-Aug. 1842, letters from the Texan Santa Fe expedition: "our horses and oxen from being stampeded" (19 July), "came near stampeding all our animals" (2 Aug. 1842 — the earliest verified *stampeding*), "stampeded no less than eighty-seven horses" (23 Aug.); Kendall's "A Stampede" extract 1844; figurative "political stampede" from 1844; first Virginia papers *Martinsburg Gazette* 8 May 1845 ("enough to have 'stampeded' a regiment of dragoons", in quotation marks) and *Staunton Spectator* 12 Feb. 1846. The date-ordered window for *stampeding* (results/lang_attest_window_stampeding_1826-1848.csv) verifies only the 2 Aug. 1842 use among the earliest 40 fuzzy hits. |
| stampedo | yes, once, as a title | *Daily National Intelligencer* 22, 23, 25 Dec. 1834 advertise the contents of the *Military and Naval Magazine of the United States* for December 1834, item "16. A Stampedo". The piece itself (vol. 4 no. 4, pp. 296-98, archive.org) is a soldier's account of "a thousand steeds ... rushing blindly and madly in a mass through the camp" at night on a marshy prairie, and ends: "The word which heads this article is a North American attempt to adopt a Spanish word into our language; but is an entire failure, so far as relates to obtaining the true Spanish sound; nevertheless, stampedo is expressive, and is in general use among those, who have witnessed similar occurrences." In 1834, then, the word was still a frontier term that needed explaining to Washington readers. This is the form Flint used in 1826 and Thomas James in 1846; it precedes Etymonline's "stampedo 1839". |
| improvised / improvise | 1830-1835 only, and only "compose or perform extempore" | *Litchfield Enquirer* 2 Dec. 1830 ("a kind of improvised account of given chapters in philosophy"); *New-Hampshire Statesman* 16 Mar. 1833 (Scott "literally improvised"); *Rhode-Island Republican* 14 Oct. 1835 ("the boy who had improvised this piece of pleasantry"). The 1803-1806 hits are OCR errors for "improving", 1825 is "Improvisatrice". A further window for 1836-1845 (results/lang_attest_window_improvised_1836-1845.csv; 8 verified of the earliest 25 hits, 1837-1840) is the same: "his improvised prayer", "improvised genius", "she improvised before she could speak", "improvised with wonderful facility" (an improvisatrice, 1839), "improvised in his honor" (1840). **No verified use through 1840 in the letters' sense "contrive [tools] from what is at hand".** |
| grizzlies | **no** (40 fuzzy hits, all *grizzly* / *grizzled*) | None in 1800-1835; a date-ordered window search for 1836-1850 (results/lang_attest_window_grizzlies_1836-1850.csv) checked the earliest 40 of 517 fuzzy hits, which reach only to Feb. 1838, and verified none. So: no verified newspaper *grizzlies* before 1838; later coverage not checked page by page (Google Books puts the take-off in the 1850s). |
| objective point | **no** (0 hits) | Richmond Daily Whig 5 Jan. 1847 (section 1). |
| reliable | ambiguous | The four pre-1812 hits are OCR-doubtful ("TO BE SOLD Reliable" 1800; "the most reliable ladies in Morgantown" 1803; "reliable ... to a suit for reparation" 1804, i.e. *liable*; "a reliable reward" 1811). No unambiguous *reliable* = trustworthy before 1822 was verified; the 1803 Portsmouth item is the only plausible one. |
| appliances | yes, mainly as a Shakespeare tag | *Daily National Intelligencer* 23 Feb. 1815, *Rhode-Island American* 11 Jan. 1820, *Massachusetts Spy* 20 Feb. 1822 — all quoting *2 Henry IV* "with all appliances and means to boot"; *Daily National Intelligencer* 1 Nov. 1822 "extraordinary appliances" (general sense). The word existed; "tools and appliances" is not what these show. |
| outfits | yes | 1801 (diplomatic "outfits"), *National Intelligencer* 12 Oct. 1807 ("vessel, outfits, and cargo"), 1813, *Constitutional Whig* (Richmond) 18 Feb. 1825. Not an anachronism. |
| tethered | yes | *Richmond Enquirer* 30 Aug. 1808, 14 Mar. 1809, 8 Sept. 1809. The Ngram flag (ratio 0.20) was a false alarm. |
| buffaloes | yes | 1816, 1820, *Edwardsville Spectator* 19 Mar. 1822. |
| worth the candle | yes | *National Intelligencer* 14 Apr. 1806 and *Richmond Enquirer* 29 Apr. 1806 ("the play is not worth the candle"), *Martinsburgh Gazette* 22 May 1812 ("the game does not appear worth the candle"), *Kentucky Gazette* 6 July 1820. |
| ho for the | yes | *Delaware Register* 14 Feb. 1829, "Then ho! for the combat" (the other hits are OCR "be for the"). |
| well nigh | yes | *Oracle Post* 18 June 1805, *Virginia Argus* (Richmond) 31 July 1805, *Alexandria Gazette* 15 Oct. 1812. |
| encounter the expense | yes | *Kentucky Gazette* 6 Nov. 1818; *Richmond Enquirer* 1 Jan. 1822; *Daily National Intelligencer* 6 Apr. and 10 June 1822. |
| personating | yes | *Alexandria Daily Gazette* 25 June 1812 ("personating the character ... of Americans"); 1813 (Matthews County, Va., "personating Englishmen"). |
| punctillio | yes | *Richmond Enquirer* 31 Jan. 1809 ("punctillios"); *Arkansas Advocate* 9 May 1832. |
| exhilerating | yes | *Providence Journal* 22 Jan. 1800, *New Hampshire Gazette* 25 Jan. 1803 and 11 Sept. 1810, *Daily National Intelligencer* 29 Oct. 1813. |
| systematize | yes | *New Hampshire Gazette* 20 May 1800; *Virginia Argus* 7 Nov. 1811; *Daily National Intelligencer* 6 Mar. 1813, 16 Dec. 1815. |

The ledger is therefore lopsided in an informative way. Every item on the "archaic" side (*punctillio,
exhilerating, well nigh, worth the candle, ho for the, encounter the expense, personating, &c.*) is
verified in American newspapers of 1800-1835, most of them in Richmond or Washington papers that a
Lynchburg writer of the 1880s could still have read or absorbed; the Ngram flags on *tethered* and
*buffalo(es)*, and the doubt about *outfits*, are shown to be false alarms; but the four words that the
screen flagged most strongly have no verified use in the same 1800-1835 newspaper corpus at all
(*stampede* family — first 1836, *stampeding* first 1842; *grizzlies* — none; *objective point* — first
1847), and *improvised* appears only from 1830 and only in the extempore-composition sense.
The archive.org full-text API (books and periodicals; top-100 hits per term, metadata date ≤ 1845,
term verified in the search highlight) adds little, because its dates are unreliable (Ballantyne's
*Dog Crusoe* of 1861 and Garland's *Ho, for the Klondike!* of 1898 both carry a metadata date of
1800) and its ranking is by relevance, not date. It confirms *punctillio* (books of 1649-1845),
*exhilerating* (1676-1845), *personating* (1806-33), *encounter the expense* (33 verified items
1812-45, e.g. *American Farmer* 25 Aug. 1820) and *outfits* (1831 Supreme Court report) as ordinary
period usage; it finds "A Stampedo" in the *Military and Naval Magazine of the United States* for
December 1834 (vol. 4 no. 4), and the racehorse name Stampedo in the General Stud Book (1807). It
returned no pre-1846 item at all for *stampede/stampeded/stampeding*, *grizzlies*, *objective point*,
*improvised*, *reliable*, *appliances*, *tethered*, *buffaloes* or *systematize* among the top 100
hits, which for the common words only reflects the ranking.

## 4. Methods

**Texts.** Primary texts as parsed from the Wikisource transcription of the Library of Congress copy
(data/primary). The 403 words that follow Morriss's quotation in `morriss_statement.txt` ("Such, in
substance, was the statement…") are the anonymous author's and were moved to the narrative; the
narrative used here is therefore 3,954 words. "Beale letters" = the three letters concatenated
(2,716 words); "letter 1" = the letter of 4 Jan. 1822 alone (2,149 words); the cipher-2 plaintext
(150 words) is too short for stylometry and is only included in the vocabulary lists.

**Control corpus.** `code/stylo_download_controls.py` fetches the raw texts (Project Gutenberg plain
text; archive.org OCR for the rest); `code/stylo_build_samples.py` strips editorial matter,
footnotes, running heads, chapter synopses and dialect passages, isolates the author's own letters
where a volume mixes editor and author (Jackson, Lee, Poe, Twain letters), and cuts up to four
2,000-word samples per work at paragraph boundaries, evenly spaced through the text so that samples
of one author are never adjacent. `data/controls/manifest.csv` lists every sample with author, title,
date written, date published, source URL, word count and genre; `data/controls/raw/download_log.tsv`
lists the source URLs of the raw files. Kennedy's *Memoirs of William Wirt* was downloaded but not
used (Wirt's letters cannot be separated from Kennedy's narrative automatically); Fowler's 1821-22
journal was downloaded but excluded from the style tests because of its phonetic spelling; the
Southern Historical Society Papers volume mixes authors and is labelled as such.

**Distances.** `code/stylo_delta.py`. Words: lower-cased alphabetic tokens (apostrophes kept, hyphens
split, "&c." kept as a token). Feature lists (the N most frequent words; the 2,000 most frequent
character 4-grams including spaces; a fixed list of 150 function words) are taken from the control
samples only, and z-scores use the control mean and sd only, so the Beale texts do not influence the
standardisation. Burrows Delta = mean |Δz|; Cosine Delta = cosine distance between z-vectors (Evert et
al. 2017); "raw cosine" = cosine on relative frequencies without standardisation. Same-author pairs =
two samples of one control author (Cooke's two works and Twain's two works count as one author each,
as do Poe's letters and tale); different-author pairs = samples of two authors. Percentiles are
empirical; likelihood ratios are ratios of Gaussian kernel-density estimates of the two distance
distributions at the observed distance (they are unstable in the extreme tails, so the length-matched
ratios of 10-25 are the ones to quote). The author-level bootstrap resamples the 27 authors with
replacement 300 times. Leave-one-out accuracy: nearest-neighbour attribution of each control sample
among the others. Period classification: mean distance to 1805-1846 samples vs 1858-1904 samples with
the sample's own author held out; the letters-only variant restricts both groups to letter-genre
samples. Classical MDS map: results/stylo_delta_mds.svg; density plot: results/stylo_delta_calibration.svg.

**Punctuation/syntax.** `code/stylo_features.py`, regular-expression heuristics applied identically
to every text: sentences split at . ! ? followed by a capital, with abbreviations protected;
infinitives = *to* + a word that cannot be a determiner/pronoun/adverb/derived noun; relativisers =
*which/who/whom/whose*; passives = form of *be* (+ adverb) + word ending in -ed/-en/-wn/-ne; negative
passives = a negator within three words of a passive; untriggered reflexives = *-self/-selves* after
*than, like, as, and, or, for, from, to, with, by, of, regarding, besides, except, but, between, upon,
on, in*. The distance in feature space is the RMS of z-differences over the rate features.

**Vocabulary dating.** `code/lang_ngram_query.py` queries the Google Books Ngram Viewer JSON endpoint
(corpus en-US-2019, smoothing 0, 1780-1900, case-insensitive) for every word type and caches the raw
series (results/lang_ngram_cache*.json); `code/lang_ngram_analysis.py` computes the ratio statistics
and the calibration; `code/lang_ngram_plot.py` draws results/lang_ngram_timeseries.svg;
`code/lang_spelling_check.py` counts the orthographic conventions; `code/lang_attest_search.py`
searches Chronicling America (loc.gov JSON API, issues dated 1800-1835) and the archive.org full-text
API for each candidate word/phrase and verifies each hit against the page OCR (the search engine
matches fuzzily, so unverified hits are discarded); `code/lang_attest_window.py TERM Y0 Y1 REGEX`
does the same for one term in a chosen year window, ordered by date, and was used to find the
earliest verified newspaper uses after 1835 (results/lang_attest_window_*.csv). Dictionary dates were
read from merriam-webster.com and etymonline.com pages fetched on 2026-09-14; the OED is paywalled and
its dates are quoted at second hand (search-engine summaries of oed.com; Mashed Radish 2015). The
HathiTrust full-text search returned HTTP 403 to this environment and could not be used; the Google
Books API returned "quota exceeded"; the Ngram endpoint's pre-1830 data are thin and include misdated
books, so a handful of non-zero pre-1822 values for *stampede* etc. were treated as noise unless a
dated page could be verified.

**Caveats.** (1) Size: 2,000-3,000-word samples give leave-one-out accuracies of only 0.6-0.7 here,
which is why the comparison is framed as a calibrated distance rather than a classifier verdict. (2)
Genre: the letters are letters and the narrative is a narrative; pronoun removal, function-word-only
features, a letters-only period test and the same-author cross-genre pairs available (Poe, Twain)
address this, and the genre penalty in this corpus is small (cross-genre different-author mean 1.025
vs 1.015 overall; cross-genre same-author 0.851 vs 0.806). (3) Editing: the letters exist only as
printed in 1885; the compositor (or the author) could have normalised spelling, punctuation and
compounding, which is why spelling evidence is weighted lightly and the semicolon difference is not
taken as evidence of different hands. (4) Influence: the narrator says he memorised the letters. (5)
Google Books frequencies are corpus artefacts (OCR, misdating, genre mix); they are used only for
screening and for the calibration, which compares like with like. (6) Nickell's article itself was
not available; his feature list is known only from secondary summaries.

## 5. Reproducibility

```
python3 code/stylo_download_controls.py      # raw texts -> data/controls/raw (+ download_log.tsv)
python3 code/stylo_build_samples.py          # samples + manifest.csv
python3 code/stylo_delta.py                  # results/stylo_delta_* (report.md, summary.json, csv, svg)
python3 code/stylo_features.py               # results/stylo_features* 
python3 code/lang_ngram_query.py results/lang_ngram_letters.csv data/primary/letter_1822-01-04.txt data/primary/letter_1822-01-05.txt data/primary/letter_1822-05-09.txt data/primary/b2_plaintext_as_printed.txt
python3 code/lang_ngram_query.py results/lang_ngram_calib_<author>.csv <2716-word control sample>   # x12
python3 code/lang_ngram_analysis.py          # results/lang_ngram_calibration.md
python3 code/lang_ngram_plot.py              # results/lang_ngram_timeseries.svg
python3 code/lang_spelling_check.py          # results/lang_spelling_conventions.md
python3 code/lang_attest_search.py           # results/lang_attestations.md (slow: verifies OCR of every hit)
python3 code/lang_attest_window.py "objective point" 1836 1848 '\bobjective point' 40   # results/lang_attest_window_*.csv
```

## 6. Sources

Primary and control texts
- The Beale Papers (Lynchburg: Virginian Book and Job Print, 1885), Wikisource transcription of the
  Library of Congress copy: https://en.wikisource.org/wiki/The_Beale_Papers (data/primary, parsed by
  sources/parse_pamphlet.py).
- Control texts: Project Gutenberg e-books 16784 (Jefferson), 43774 (Pike), 16565 (Biddle/Lewis &
  Clark), 44205 (Gregg), 46110 (Pattie), 62018 (Fowler), 2323 (Lee), 19831 (Davis), 26725 (Page),
  10234 (Cable), 51211 (Eggleston), 245 (Twain), 3195 (Twain letters), 10692 (Cooke, Life of Lee),
  2147 (Poe, Works vol. 1); archive.org items selectionsfrommi02bagbrich (Bagby 1885),
  virginiaahistor02cookgoog (Cooke 1883), correspondenceof0003jack (Jackson, Bassett ed. 1928),
  memjohnquincy05adamrich (J. Q. Adams, Memoirs vol. 5), recollectionsofl00flin_0 (Flint 1826),
  lettersofjohnran00rand (Randolph 1834), threeyearsamongi00jame (Thomas James 1846/1916),
  sketchesrecollec00cabe (Cabell 1858), lynchburgitspeop00chri (Christian 1900),
  sketchbookoflync00poll (Pollock 1887), accountofexpedit02jame (Edwin James 1823),
  lettersfromsouth00paul (Paulding 1817), lettersfromwestc00hall (Hall 1828),
  lifeandletterse00harrgoog (Harrison, Life and Letters of Poe, 1903), southernhistoric12sout (SHSP
  vol. 12, 1884), touronprairies00irvi (Irving 1835). Full URLs in data/controls/manifest.csv.

Nickell, Pival, Poundstone (secondary, since JSTOR 4248566 is paywalled)
- Joe Nickell, "Discovered: The Secret of Beale's Treasure", VMHB 90:3 (July 1982) 310-324,
  https://www.jstor.org/stable/4248566 (not consulted directly).
- Wikipedia, "Beale ciphers" (summary of Nickell: anachronisms; "punctuation, relative clauses,
  infinitives, conjunctives"; Poundstone's stylometric exclusion of Poe, citing *Big Secrets* 1983,
  p. 133): https://en.wikipedia.org/wiki/Beale_ciphers
- Abbeville Institute, "The Mystery of the Beale Treasure" (quotes Pival: "misuse of reflexive
  pronouns and frequent use of negative passive phrases such as 'never to be told' and 'never to be
  realized'"): https://www.abbevilleinstitute.org/the-mystery-of-the-beale-treasure/
- Damn Interesting, "89, 263, 201, 500, 337, 480" (Pival: "nearly identical average sentence length,
  as well as very similar proportions of the words the, and, and of"; Nickell: stampede/improvise "not
  known ... before 1840", stampeding "unattested before 1883"; Morriss took the Washington Hotel only
  at the end of 1823): https://www.damninteresting.com/89-263-201-500-337-480/
- Joe Nickell, "Is the Beale Treasure Haunted?", Center for Inquiry blog (his own summary: linguistic
  analysis "pointed to an obvious suspect, one James Ward"):
  https://centerforinquiry.org/blog/is-the-beale-treasure-haunted/

Dictionaries and attestations
- Merriam-Webster: stampede (noun 1828, verb 1838; American Spanish *estampida*)
  https://www.merriam-webster.com/dictionary/stampede; stampedo (archaic, from *estampido*)
  https://www.merriam-webster.com/dictionary/stampedo; improvise (1788)
  https://www.merriam-webster.com/dictionary/improvise; grizzly (adj. 1594; noun "early 19th
  century") https://www.merriam-webster.com/dictionary/grizzly; reliable (adj. 1569, noun 1852)
  https://www.merriam-webster.com/dictionary/reliable; objective (adj. 1645, noun 1835)
  https://www.merriam-webster.com/dictionary/objective; outfit (noun 1748), punctilio (1596),
  personate (1591), buffalo (1562), appliance (1555), locality (1538), shipment (1783), bracing (1752).
- Etymonline: stampede (verb 1823; noun 1836, variants stampiado 1828, stampedo 1839)
  https://www.etymonline.com/word/stampede; improvise (1808) https://www.etymonline.com/word/improvise;
  grizzly ("grizzly bear ... recorded by 1806") https://www.etymonline.com/word/grizzly; reliable
  ("not common before 1850, and often execrated thereafter in Britain as an Americanism")
  https://www.etymonline.com/word/reliable; objective ("objective point" 1852, "goal" 1881)
  https://www.etymonline.com/word/objective.
- OED (paywalled) as quoted: stampede n., earliest evidence 1828 ("stampiado"), search-engine summary
  of https://www.oed.com/dictionary/stampede_n; the 1823 verb citation from the Austin Papers and
  Flint's 1826 "stompado" per Mashed Radish, "Stampede" (29 Sept. 2015),
  https://mashedradish.com/2015/09/29/stampede/
- The Austin Papers, ed. E. C. Barker, Annual Report of the American Historical Association for the
  Year 1919, vol. II (Washington, 1924), part 1, p. 699: "Deposition concerning loss of horses"
  (following J. Cable to Austin, Natchitoches, 14 Oct. 1823; events of Aug. 1822 at La Bahía): "the
  Cavellada Stampeded", "the stampede etc.", "the Cavellada that stampeded"; also Spanish "estampida"
  in a document of 12 Jan. 1812 (p. c. 220). OCR text:
  https://archive.org/details/sim_american-historical-association-annual-report_1919_2
- Timothy Flint, *Francis Berrian, or the Mexican Patriot* (Boston, 1826), the wild-horse passage
  "took what the Spanish call the 'stompado'", as reprinted in the Oxford Observer (Paris, Maine),
  2 Nov. 1826 (https://archive.org/details/SERIAL_8.1826.11.02), Catawba Journal 7 Nov. 1826
  (https://archive.org/details/per_journal-1833_catawba-journal_1826-11-07_3_106), Saturday Evening
  Post 25 Nov. 1826 (https://archive.org/details/sim_saturday-evening-post_1826-11-25_5_47_0), The
  Casket, Jan. 1827, and The Class Book of American Literature (1826,
  https://archive.org/details/classbookofameri00frosrich); found with the archive.org full-text API
  (https://be-api.us.archive.org/fts/v1/search?q=%22stompado%22).
- Josiah Gregg, Commerce of the Prairies (1844), Thwaites ed. (Gutenberg 44205): "a stampede", "who
  _stampeded_ and carried away" (italics in the original), pp. 168 and 230 of vol. II.
- Thomas James, Three Years Among the Indians and Mexicans (1846; 1916 reprint,
  https://archive.org/details/threeyearsamongi00jame): "A stampedo", "gave my drove a stampede",
  "lost by flies and stampedos".
- M. E. W. Sherwood, Manners and Social Usages (1887 ed., https://archive.org/details/mannerssocialus00shergoog):
  "such a stampede of clocks and lamps" — the source of Wikipedia's misdated "1834" citation.
- "Objective point", earliest verified newspaper uses (all Mexican-War commentary; OCR checked by
  code/lang_attest_window.py, results/lang_attest_window_objective_point_1836-1848.csv): Richmond
  Daily Whig, 5 Jan. 1847, p. 2, https://www.loc.gov/resource/sn84024656/1847-01-05/ed-1/?sp=2 ;
  Daily Union (Washington, D.C.), 9 Jan. 1847, p. 2,
  https://www.loc.gov/resource/sn82003410/1847-01-09/ed-1/?sp=2 ; Staunton Spectator, 14 Jan. 1847,
  p. 2, https://www.loc.gov/resource/sn84024719/1847-01-14/ed-1/?sp=2 ; Alexandria Gazette, 15 Jan.
  1847, p. 2, https://www.loc.gov/resource/sn85025007/1847-01-15/ed-1/?sp=2 ; Boon's Lick Times
  (Fayette, Mo.), 23 Jan. 1847, p. 1, https://www.loc.gov/resource/sn83016957/1847-01-23/ed-1/?sp=1 .
  No verified use in Chronicling America for 1800-1846 (search 1836-1848 returned only these 1847
  items; the 1800-1835 search returned nothing).
- Chronicling America search API (loc.gov): https://www.loc.gov/collections/chronicling-america/?q=...&dates=1800/1835&fo=json ;
  the two 1824 "stampede" hits (American Watchman 1 June 1824; Daily National Intelligencer 26 May
  1824) read "stamped, or painted" in a tariff act.
- Chronicling America attestations verified against the page OCR by code/lang_attest_search.py and
  code/lang_attest_window.py (results/lang_attestations.csv, results/lang_attest_window_*.csv):
  stampedo — Daily National Intelligencer 22 Dec. 1834 p. 1 (contents of the Military and Naval
  Magazine, Dec. 1834), https://www.loc.gov/resource/sn83026172/1834-12-22/ed-1/?sp=1 ;
  stampede — Arkansas Advocate 12 Aug. 1836 p. 2, https://www.loc.gov/resource/sn87062070/1836-08-12/ed-1/?sp=2 ;
  New-Hampshire Statesman 23 Mar. 1839 p. 3, https://www.loc.gov/resource/sn84022931/1839-03-23/ed-1/?sp=3 ;
  stampeded / stampeding — Southern Argus (Columbus, Miss.) 19 July 1842 p. 1,
  https://www.loc.gov/resource/sn83016884/1842-07-19/ed-1/?sp=1 ; 2 Aug. 1842 p. 1,
  https://www.loc.gov/resource/sn83016884/1842-08-02/ed-1/?sp=1 ; 23 Aug. 1842 p. 1,
  https://www.loc.gov/resource/sn83016884/1842-08-23/ed-1/?sp=1 ;
  improvised — Litchfield Enquirer 2 Dec. 1830 p. 1, https://www.loc.gov/resource/sn84020071/1830-12-02/ed-1/?sp=1 ;
  New-Hampshire Statesman 16 Mar. 1833 p. 2, https://www.loc.gov/resource/sn84022931/1833-03-16/ed-1/?sp=2 ;
  Rhode-Island Republican 14 Oct. 1835 p. 4, https://www.loc.gov/resource/sn83025561/1835-10-14/ed-1/?sp=4 ;
  appliances — Daily National Intelligencer 23 Feb. 1815 p. 2, https://www.loc.gov/resource/sn83026172/1815-02-23/ed-1/?sp=2
  and 1 Nov. 1822 p. 2, https://www.loc.gov/resource/sn83026172/1822-11-01/ed-1/?sp=2 ;
  reliable (OCR-doubtful) — United States Oracle (Portsmouth, N.H.) 30 Apr. 1803 p. 4,
  https://www.loc.gov/resource/sn83025545/1803-04-30/ed-1/?sp=4 ;
  punctillio — Enquirer (Richmond) 31 Jan. 1809 p. 3, https://www.loc.gov/resource/sn84024736/1809-01-31/ed-1/?sp=3 ;
  exhilerating — Providence Journal 22 Jan. 1800 p. 1, https://www.loc.gov/resource/sn83021629/1800-01-22/ed-1/?sp=1 ;
  well nigh — Virginia Argus (Richmond) 31 July 1805 p. 3, https://www.loc.gov/resource/sn84024710/1805-07-31/ed-1/?sp=3 ;
  worth the candle — Enquirer (Richmond) 29 Apr. 1806 p. 1, https://www.loc.gov/resource/sn84024736/1806-04-29/ed-1/?sp=1 ;
  ho for the — Delaware Register 14 Feb. 1829 p. 8, https://www.loc.gov/resource/sn84020593/1829-02-14/ed-1/?sp=8 ;
  encounter the expense — Richmond Enquirer 1 Jan. 1822 p. 3, https://www.loc.gov/resource/sn84024735/1822-01-01/ed-1/?sp=3 ;
  personating — Alexandria Daily Gazette 25 June 1812 p. 2, https://www.loc.gov/resource/sn84024013/1812-06-25/ed-1/?sp=2 ;
  tethered — Enquirer (Richmond) 30 Aug. 1808 p. 3, https://www.loc.gov/resource/sn84024736/1808-08-30/ed-1/?sp=3 ;
  outfits — National Intelligencer 12 Oct. 1807 p. 1, https://www.loc.gov/resource/sn83045242/1807-10-12/ed-1/?sp=1 ;
  buffaloes — Edwardsville Spectator 19 Mar. 1822 p. 2, https://www.loc.gov/resource/sn82015374/1822-03-19/ed-1/?sp=2 ;
  systematize — Virginia Argus 7 Nov. 1811 p. 1, https://www.loc.gov/resource/sn84024710/1811-11-07/ed-1/?sp=1 .
- Military and Naval Magazine of the United States (Washington), vol. 4 no. 4 (Dec. 1834), "A
  Stampedo", pp. 296-98 (quotation "The word which heads this article is a North American attempt to
  adopt a Spanish word into our language ... stampedo is expressive, and is in general use among those,
  who have witnessed similar occurrences"): OCR text
  https://archive.org/download/sim_military-and-naval-magazine-of-the-united-states_1834-12_4_4/sim_military-and-naval-magazine-of-the-united-states_1834-12_4_4_djvu.txt
  (item https://archive.org/details/sim_military-and-naval-magazine-of-the-united-states_1834-12_4_4 ;
  found with the archive.org full-text API; also in the volume index,
  https://archive.org/details/sim_military-and-naval-magazine-of-the-united-states_1834_4_index).
- Google Books Ngram Viewer JSON endpoint, corpus en-US-2019:
  https://books.google.com/ngrams/json?content=...&year_start=1780&year_end=1900&corpus=en-US-2019&smoothing=0&case_insensitive=true
- Evert, Proisl, Jannidis et al., "Understanding and explaining Delta measures for authorship
  attribution", Digital Scholarship in the Humanities 32 suppl. 2 (2017) ii4-ii16 (Cosine Delta).

## Addendum: the likelihood ratios as quoted in the report

The report quotes the length-matched likelihood ratios as the full range across the five reported measures, 2.7
(Burrows Delta, 300 MFW, 2,000 words; the measure with the best leave-one-out accuracy on the controls, 69 %) to 25
(pronoun-free Cosine Delta and character 4-grams), rather than "10 to 25", and gives the share of different-author
pairs as close as the Beale pair as 0.0 to 4.9 % depending on the measure (results/stylo_delta_summary.json). With even
prior odds that is P(same hand) 0.73 to 0.96. The same-topic control pairs are farther apart than the Beale pair on
average (means 0.85 to 0.99 against 0.70) but one individual sample pair, Christian vs Pollock, is closer (0.628). The
control rate for the "not/never to be + participle" construction is 0.03 per thousand words (ctrl_mean 0.0287 in
results/stylo_features_summary.json), not 0.05.