8. What this report adds, and where it falls short
No new evidence is presented here: no document was found that was not already in print. The documentary case on both sides was assembled between 1946 and 2008. What is new is analysis of old evidence, mainly the added tests of the 2019 design. A second implementation of those tests, written independently, reproduces the attribution, finds the register effect in the same direction, and finds it smaller (section 4).
Established before this report
- Bullard (1946): the route of the request, the Transcript text, the forged facsimiles, the chains behind the Morley, Brownell and Page stories, Hay's letters to Chandler (1904) and Langford (1903), and the argument from Lincoln's other writing.
- Nickell and Pival (1989): the parallels in Lincoln's own letters ("tender", "our Father in Heaven", "sacredness of your sorrow"), the alliteration, and the syntactic markers ("should", the active voice) that favour Lincoln.
- Burlingame (1995, 1999): the five reported claims assembled, "beguile" and the other Hay-marked words, the Spencer letter, and the scrapbooks at Brown and the Library of Congress.
- Emerson (2006, 2008): Robert Lincoln's Markens letters, with Hay's statement that he knew nothing of the letter's source; the counter-argument on the scrapbooks; the fifth sentence of the Herndon letter.
- Grieve and colleagues (2019): n-gram tracing as a method for very short texts, its validation on 1,662 known texts, and the attribution of the letter to Hay under their design.
- Grieve and colleagues (2016 presentation): the caution, in their own words, that the method's sensitivity to register variation still needed testing.
- Thomas Proisl and colleagues, computational linguists in Germany (2018): that n-gram tracing is "highly sensitive to the choice of authors and texts in the comparison corpus".
Added here
- The first independent replication of the 2019 attribution, with the method re-implemented from the paper's description and validated on the same kinds of texts.
- Three tests of that design's sensitivity, which its authors had named as untested: period (Lincoln's wartime own-hand texts of 100 to 200 words go to Hay 14 percent of the time under a pre-1860 Lincoln corpus), register (30 of 63 register-matched Lincoln pieces go to Hay in our implementation and 19 of 63 in the second, among them in both the Gettysburg Address and the Ellsworth condolence), and edition (print versions of 95 Lincoln texts sit nearer Hay than their autograph versions at every string length).
- A corpus design built from Basler's source codes, so that Lincoln's side holds only what survives in his hand from 1860 to 1865, and a register-matched validation set of Hay's own war prose of 1861.
- The test that matters most: under the own-hand design 18 of the 19 Basler items in Hay's hand, three of Hay's four known letters for Lincoln's signature, and Hay's own condolence of 1864 are called Lincoln by the vote, and the classifiers and Delta call all five of the known Hay texts Lincoln. The methods cannot detect the hypothesis at issue, so a result on the letter under them is not evidence for either man.
- A word and phrase table across four corpora that places the individual examples of Burlingame and Nickell in their base rates.
- The documentary items weighed as explicit likelihood ratios, with the sensitivity of the result to each published beside the verdict.
- A count of Basler's items for every week of the autumn of 1864, which puts a number on Bullard's answer to the "too busy" argument.
- The observation that the Phillips copy of 21 November 1864, kept beside the Bixby clipping in Hay's scrapbook, is in a hand that is not Hay's. We did not find this pointed out before.
Since 2019
We found no published replication, re-analysis or rebuttal of the 2019 attribution between its appearance and September 2026, in the journals, in the 41 works citing it that the Semantic Scholar database lists, or in the Lincoln periodicals we could reach. The only independent test of the method is Proisl and colleagues (2018), on literary texts, published before the journal version of the study. A new book, Jan Cigliano Hartman's Lincoln's Speechwriter (2026), argues for a large Hay hand in Lincoln's prose generally; the Lincoln scholar Harold Holzer's review of it (on the history website Emerging Civil War, 23 July 2026) calls the Bixby question open and names the scrapbook as the principal argument for Hay. We did not read the book. The journal and book databases Project MUSE, JSTOR, HathiTrust and Google Books could not be searched, so the search is not complete.
Where this report falls short
- No archive was visited and no new document was found. The Hay Papers, the Morley papers and the Bixby family papers were not searched.
- Basler's volumes II, III, V and VI (1848 to 1863) were lending-only and are absent from the Lincoln corpus. The own-hand design covers 1860 to 1865 from volumes IV, VII and VIII, and the pre-1860 design uses volume I and the Lapsley edition. The McCullough condolence of 1862 enters the register set from a printed text.
- Burlingame's 2000 chapter, the fullest statement of the case for Hay, was not available and is cited through Emerson and Burlingame's own later summaries; Edward Steers's Lincoln Legends (2007), which discusses the letter, was not read.
- Hay's letter to Spencer of 25 November 1864 is quoted from Burlingame's quotations of it, which differ in one word between his 1999 article and his later book.
- The register-matched Hay set is journalism and memoir of 1861, because Hay's surviving condolence writing from the 1860s is a single letter of 55 words. A true test of the hypothesis would need Hay condolences that do not exist in print.
- The impostors method (a check that asks whether a candidate still wins when many unrelated writers are added to the comparison), planned as a further check for section 4, did not finish and is not reported.
- The added tests exist in two implementations, written independently of each other. They agree on the direction of the register effect and differ on its rate, and section 4 traces the difference as far as we could. Neither used the paper's own corpus, which was not obtainable, and no one else has rerun them.
- The likelihood ratios in section 5 are judgments. The sensitivity table shows how far the result moves when any one of them is changed; a reader who disagrees with several at once can end up on the other side.