The Casket Letters
Eight letters and a run of sonnets, said to be in Mary Queen of Scots' hand, were used to depose her. The originals are gone. Can what survives still tell us who wrote them?
In June 1567 Mary's enemies said they had found a silver casket holding letters from her to the Earl of Bothwell. Read as genuine, the letters tie her to the February 1567 murder of her husband, Lord Darnley. They were shown to Elizabeth I's commissioners at York and Westminster in 1568. The originals disappeared in the 1580s. Historians have argued for over 450 years about whether they are genuine, forged, or a mix.
This page is the public record of Stratah's attempt at a narrow, testable part of that question. Every finding, every dead end and every change of plan is logged below and stays visible.
To the best of our knowledge as of 2026-10-01. First result: the surviving French letters lean toward Mary's phrasing over Catherine de Médicis's (claim C-06, proposed, not yet reviewed). A genre confound is still open. An earlier test design (DT-001) was retired before running because the French text is too short for it.
The question
Original question (too broad): Are the Casket Letters genuine?
Narrowed open question question: Do the casket texts that survive in French match the writing habits of Mary's authenticated French letters more closely than they match other French writers of her time?
Why it was narrowed: the most important letter, the long "Glasgow letter" (Letter II), survives only in Scots, English and Latin versions. Style tests on translations are unreliable (see finding F2), so that letter cannot be tested this way. The narrowed question is smaller, but an answer to it would be real evidence instead of another opinion.
Claims
Each claim shows evidential weight (cool) and adoption weight (warm) separately. Adoption is information, not evidence. anchor_checked: no means we have only read secondary descriptions so far, not the primary document itself.
Primary text: the forged postscript survives in the UK National Archives and is described consistently by the NSA history and archive sources.
Standard account in every modern history of the Babington Plot.
No original has been traced. They were returned to the Earl of Morton in 1571 and vanished some years after his 1581 execution.
Universal.
Depth: who destroyed them?
Several sources say Mary's son, James VI, is "believed" to have destroyed them in 1584. That attribution is weaker than the loss itself and is held separately as a proposed claim, not merged into C-02.
Argued by T. F. Henderson (1889–90), reported by the 1911 Britannica. Checkable by comparing the texts side by side.
Standard among specialists; little known outside them.
Depth: why this matters for the test
If C-03 holds, the printed French texts must be excluded from the test entirely. Only the French copies found among William Cecil's papers (Hatfield) can stand in for the lost originals.
No originals. Content and dating arguments cut both ways. No independent test has been applied yet.
Minority among modern biographers. National Museums Scotland says it is "widely thought" they were doctored.
Absence from two web searches only. This is weak. A specialist could know of unpublished or non-indexed work.
Not a claim anyone holds yet.
Revision 2026-10-01 (finding F7)
Narrowed. Manual comparisons do exist: Henderson reports that Harry Bresslau found the French opening of Letter II "corresponds very closely in phrase and structure" with a genuine letter in Labanoff, and that Bernard Sepp argued from idioms in the translations. What remains unfound is a systematic, computed comparison across all surviving French text.
Our own computed result (DT-002), calibrated at about 96% on held-out letters. Weakened by a genre confound and by OCR copies. Not yet reviewed by anyone outside Stratah.
New today. Agrees in direction with Bresslau (1882), but that is agreement, not confirmation.
Positions on authenticity
The three live positions. The weights are kept in separate columns on purpose.
| Position | Held by (examples) | Evidential basis | Adoption |
|---|---|---|---|
| Genuine | Robertson and Hume (18th c.); T. F. Henderson (1889); Jenny Wormald | Most of Elizabeth's commissioners accepted them after comparing the handwriting with Mary's. Wormald argues that the lords' reluctance to produce them, and their destruction, imply real evidence. Both are inferences, not tests. | Low to moderate |
| Forged | Adam Blackwood (16th c.); Walter Goodall (1754); William Tytler | Strong motive, a broken chain of custody, no signatures, addresses or dates. Motive and opportunity show a forgery was possible, not that one happened. | Low to moderate |
| Composite (real letters spliced with forged or altered text) | Harry Bresslau (1882; Letters III–VI genuine, Letter II partly forged); Antonia Fraser (1969); John Guy (2004) | Guy estimates about 1,500–1,800 words of Letter II are likely genuine and 1,000–1,200 likely interpolated, from internal inconsistencies. This is careful reading, but it is still judgment, not measurement. | Moderate; the most common modern view |
Divergence note. The composite view has the most adoption, but none of the three positions rests on an independent test. All the evidential weight is interpretive. That gap is why this question is an open question.
Feasibility check, 2026-10-01
Before building anything, we checked whether the test could actually be done. Four findings changed the plan.
oracle
The 2023 letters can't be the main baseline yet
The 2023 paper by Lasry, Biermann and Tomokiyo is open access, but it prints one full letter plus summaries of the rest. In 2024, historian Estelle Paranque said only three historians had full access to the letters.
Mary's other French letters are fully public: Labanoff's seven-volume 1844 edition is digitised on the Internet Archive, HathiTrust and Gallica.
→ Effect: Labanoff becomes the main authenticated corpus. The 2023 letters become a later held-out check if the team grants access.
method
Style tests on translations don't work
A 2024 study tested standard authorship methods on translated texts and found they did not correctly attribute them. A translator's habits overwrite the author's.
→ Effect: Letter II (Scots, English and Latin only) is excluded from the test. Only French-language copies are tested.
sources
Not all French texts are equal
The French printed in the 1570s appears to be a retranslation (claim C-03). The French copies among Cecil's papers are the closest surviving thing to the originals. Henderson (1889) and Goodall (1754) both print the casket texts and are public domain.
→ Effect: the test uses only the Cecil/Hatfield French copies. Which letters and sonnets exist in that form is still to be listed (gap G-01).
prior work
Nobody seems to have run this test
Two searches found no computational style study of the casket texts (claim C-05). This is weak evidence of absence, so it stays a gap until a wider search and a specialist confirm it.
→ Effect: proceed, but ask before claiming novelty. Revised by F7.
sources
Only four letters survive as French copies
Henderson (1889, p. 68): "The copies of the original French versions of the letters now known to exist are Nos. 3 and 5 in the Record Office, and Nos. 4 and 6 at Hatfield." Letters 7 and 8 exist only in print. Henderson calls No. 6 the only one "in itself fatally compromising," so the testable set does include a damaging letter.
→ Effect: test set fixed at Letters III, IV, V and VI, plus the two Letter II fragments for phrase checks only.
method
The French text is too short for the planned test
A line-by-line count of Henderson's appendix gives roughly 630, 500, 240 and 400 words for Letters III–VI, about 1,750 in all (±20%, from OCR). Eder (2010) found word-frequency attribution "simply disastrous" below about 3,000 words for novels, with about 2,500 for Latin prose at best.
→ Effect: DT-001 would almost certainly fail its own calibration bar. Retired before running, with this reason. Replaced by DT-002.
prior work
Nineteenth-century scholars already compared phrases by hand
Henderson reports Bresslau's close match between Letter II's French opening and a genuine Labanoff letter, and Sepp's arguments from idiom. Henderson also cites the Hatfield calendar editors: none of the letters "can be used as direct evidence against Mary, and some of them have been suspiciously manipulated."
→ Effect: the useful new step is to make that phrase comparison systematic and repeatable, with a control baseline Bresslau did not have.
pilot
Pilot: the method finds a real parallel on its own
We searched about 77,000 overlapping passages in Labanoff volumes 1–2 (1542–1569) for the French opening of Letter II, "Estant party du lieu ou j'avois laissé mon cœur…", without pointing the search at any letter.
The third-ranked passage is Mary to Elizabeth, Ripon, 27 January 1569: "estant contrainte partir du lieu où j'ay résiday jusque à présent…". Same opening construction. The two ranked above it were noise: a garbled scan and a modern French translation of a letter Mary wrote in another language.
Two limits. Henderson does not name the letter Bresslau used, so we cannot confirm this is his parallel. And the Ripon letter postdates the casket hearings (December 1568), so it can show a habit of phrase, not a source a forger copied from.
→ Effect: positive control partly passed. Before scaling, translated and editorial text must be filtered out of Labanoff. Scans are clean enough to proceed.
prior work
The control problem was raised in the 1880s
Henderson (pp. 69–70): Bresslau made "an exhaustive comparison of the phrases and turns of style of the four French letters" with Labanoff and concluded Mary wrote them. But "it has been shown that similar peculiarities of style to those adduced by him may be found in the letters of Catherine de Medici."
→ Effect: the comparison with Mary has been done; what was missing then and now is a measured control. Catherine de Médicis's letters become the required control corpus, not an optional one. This is where Stratah can add something new.
Decision: go, narrowed. The test is doable with public sources, but it can only speak to the French-language casket texts, not the Glasgow letter. A clean result would be new evidence. It would not settle Mary's guilt, and this page will not claim it does.
Discriminating test DT-002 (current)
Question: Do the surviving French casket texts share long phrases with Mary's authenticated letters more often than chance and ordinary period French would explain?
This is phrase matching (text-reuse detection), not style counting. It works on short texts, and it speaks directly to the composite theory: a forger splicing Mary's real letters, or Mary writing in her own habitual phrases, would both leave shared phrases. A total absence would count against both.
- Mary corpus. All French letters in Labanoff, OCR from the public scans, spelling normalised (the copies were made by clerks).
- Control corpus. French letters of the 1560s by other writers, of similar size, so we know how often unrelated writers share phrases by chance.
- Match. Find every shared run of 4 or more content-bearing words between each casket text and each corpus. Discard stock formulas (greetings, closings) using the control corpus.
- Pre-set bar. A casket text counts as "phrase-linked to Mary" only if its rate of shared phrases with Mary's corpus is at least 3 times its rate with the control corpus, and each match is listed for human review.
- Positive control. The method must rediscover Bresslau's Letter II parallel without being pointed at it. If it can't, it stops and the failure is published.
- Publish code, corpora lists, every match and every non-match. Outside review before any claim moves from proposed.
Possible outcomes: phrase-linked (consistent with genuine or composite, cannot separate them); not linked (counts against both, unless the corpus is too thin); inconclusive (OCR or corpus too poor, logged as a gap).
Result, 2026-10-01
Method: each casket text's three-word phrases (spelling normalised, phrases made only of common words like "de la" ignored) were checked against equal-sized samples of Mary's French (Labanoff vols 1–2, about 31,800 words of old-spelling text) and Catherine de Médicis's (La Ferrière vols 2–3). The ratio is Mary-share ÷ Catherine-share; above 1 leans Mary. Each was repeated over 30 random samples.
| Text | French words | Ratio (median) | 90% range | Reading |
|---|---|---|---|---|
| Letter III | 657 | 1.59 | 1.25–1.78 | Leans Mary |
| Letter IV | 443 | 1.49 | 1.23–1.88 | Leans Mary |
| Letter V | 226 | 1.55 | 1.10–2.66 | Leans Mary, wide range (short text) |
| Letter VI | 387 | 1.35 | 0.94–1.64 | Leans Mary, range touches 1 |
| Letter II opening | 29 | 0.01 | 0.01–0.03 | Too short to read; ignored |
| Calibration: Mary's own letters, held out | ~420 each | 1.44 | 1.00–2.44 | 57 of 60 placed correctly |
| Calibration: Catherine's letters, held out | ~420 each | 0.64 | 0.46–0.88 | 58 of 60 placed correctly |
design error
Our pre-set bar of 3× was wrong, and we say so
DT-002 said a text is "phrase-linked to Mary" only at a ratio of 3 or more. Under that rule, no casket letter qualifies. But Mary's own genuine letters reach 3 only once in 60 tries, so the bar could never have worked. The rule that the calibration actually validated is "above 1".
→ Effect: both readings are published. Under the pre-set rule: no letter linked. Under the calibrated rule: all four lean Mary. Future bars will be set from a calibration run, not guessed.
confound
The control may differ by genre, not by author
Catherine's published letters are mostly official dispatches. The casket letters are intimate. The shared phrases driving the result include "toute ma vie", "de ma vie", "en ce monde" and "ce porteur vous dira": personal-letter language. Mary's corpus has more personal letters than Catherine's, so the lean could partly be "personal letter" rather than "Mary".
→ Effect: C-06 stays proposed. The next control must be personal letters by another French-writing woman of the period (gap G-05).
What this does and doesn't show. The surviving French letters read like Mary's letters more than like Catherine's, about as strongly as Mary's own letters do. That fits "genuine" and fits "composite" (a forger working from her real letters). It does not show guilt, and the genre gap could still explain part of it.
Run files: README · extract.py · pilot.py · compare.py · boot.py · casket_fr.json · results
Discriminating test DT-001 (retired before running)
Kept for the record. Retired 2026-10-01 because of finding F6: the French text is far below the length this method needs. No data was collected under it.
Rules are fixed here before any result exists, so the result can't shape the method. Any change after a run is logged as a change, with the reason.
- Build Mary's corpus. Her French letters from Labanoff, 1560–1568, the years closest to the casket letters. Note which are autograph and which are secretary-written.
- Build controls. French letters by other writers of the period, including women in her circle where texts exist. For the sonnets, add French court poetry of the period, since a sonnet and a letter differ in style regardless of author.
- Measure common words, not spelling. The copies were made by clerks who changed spelling, so the test uses function words (de, que, et, je, vous…) and sentence habits that copyists rarely alter.
- Calibrate first. On samples of the same length as the casket texts, the method must assign held-out Mary letters to Mary and control letters to others at least 90% of the time. If it fails, the test stops and the failure is published.
- Then score the casket texts, one at a time, with uncertainty ranges.
- Publish everything: corpus lists, code, raw scores and the calibration result, so anyone can rerun it.
- Outside review. A named specialist reviews before any claim moves from proposed.
What each result would and would not show
Matches Mary
Consistent with genuine or composite. A forger splicing Mary's real letters would also pass. Cannot prove guilt.
Does not match
Points away from Mary's authorship of that text, unless the mismatch is explained by genre or the copyist. Calibration in step 4 is what makes this result count.
Inconclusive
Texts too short or calibration fails. Logged as a searched_gap with the exact reason. Still a published result.
Conspiracy overlay
The popular claim: "Mary was framed by the state." Part of it is documented. Part of it is not. Stratah keeps the two apart.
Documented (1586)
English intelligence forged text in Mary's correspondence to trap her and her allies (C-01).
Unproven (1567–68)
The Casket Letters were forged by the Scottish lords who deposed her (C-04).
A proven forgery in 1586, by different people in another country, shows that forging Mary's letters was possible and was done. It adds no evidence that the 1567 letters were forged. Treating it as if it did is exactly the move a narrative makes when it borrows weight from a true story.
Open gaps
Excavation log
Append-only. Entries are never edited or removed; corrections are added as new entries.
Open Question Q-001 opened. Chosen over the Monteagle letter, the Zinoviev letter, Linear Elamite and the Copper Scroll, because it is old, has a conspiracy overlay with a documented core, and gained a new comparison corpus in 2023.
2023 plaintexts not fully public. Main baseline switched to Labanoff (1844).
Translated texts are unreliable for style tests. Letter II removed from scope.
Printed 1570s French likely a retranslation (C-03). Test limited to the Cecil/Hatfield French copies.
No prior computational study found in two searches. Logged as searched_gap C-05, not as a novelty claim.
DT-001 fixed with a 90% calibration bar before any data was collected.
Read Henderson's appendix directly. French copies survive for Letters III–VI only. G-01 closed.
Counted the French: about 1,750 words. Below the published floor for word-frequency methods. G-02 closed. DT-001 retired before any run.
Found 19th-century manual comparisons (Bresslau, Sepp). C-05 narrowed by a revision note.
DT-002 (phrase matching against a control baseline, with Bresslau's parallel as a positive control) set before any data was collected.
DT-002 pilot on Labanoff vols 1–2. Ripon letter of 27 Jan 1569 surfaced unprompted at rank 3 with the same opening construction as Letter II. Two noise hits above it traced to OCR and a translated letter. Positive control: partial pass.
Henderson reports Bresslau's 1882 comparison and the objection that Catherine de Médicis's letters show the same traits. Control corpus made mandatory.
DT-002 run. Calibration 115/120 correct. Letters III–VI all lean Mary (median ratios 1.35–1.59). Claim C-06 filed as proposed. Code and outputs published with this page.
Pre-set 3× bar found unusable (fails Mary's own letters); reported both ways. Genre confound identified; gap G-05 opened.
Added A-08 (Eikon Basilike) as next candidate: long enough for style tests, which the Casket Letters were not. Added A-09 (Gowrie / Logan letters), parked.
Areas-to-explore list started. Weizmann acetone / Balfour Declaration (A-02) added as the next candidate, continuing the existing WWI chemistry dig.
Challenge a finding
Every claim, finding and step on this page can be disputed. To challenge one, open an issue on the Stratah repository on GitHub (issues) naming the ID (for example C-03 or F2) and the source that anchors your challenge. A challenge with an anchor is logged here whether it is accepted or not.
Areas to explore
The running queue of candidate open questions. A candidate moves up when it has an oracle and open sources. Status changes are logged, not overwritten.
| ID | Area | Conspiracy overlay | Oracle (what could check an answer) | Status |
|---|---|---|---|---|
| A-01 | Casket Letters, 1567 | "Mary was framed" | Mary's authenticated French letters | active · Q-001 |
| A-02 | Weizmann's acetone and the Balfour Declaration, 1915–1917. Continues the existing dig: WWI ammunition chemistry → modern Israel → WWII. | "The Balfour Declaration was a reward for Weizmann's cordite chemistry," popularised by Lloyd George's war memoirs | War Cabinet minutes, Ministry of Munitions records, Weizmann's correspondence: does the acetone work appear in the decision record? | next candidate |
| A-03 | Monteagle letter, 1605 | "Cecil knew of the Gunpowder Plot early" | The original survives (TNA SP 14/216/2); handwriting exemplars of the suspects | needs a paleographer |
| A-04 | Zinoviev letter, 1924 | "Intelligence services rigged an election" | Weak: no original letter exists, only copies | parked · no oracle |
| A-05 | Linear Elamite, c. 2000 BCE | None | New inscriptions tested against the 2022 sign values | needs a specialist |
| A-06 | Copper Scroll, 1st c. CE | "Temple treasure is still buried" | None known | parked · no oracle |
| A-08 | Eikon Basilike, 1649 (now open question Q-002 (formerly bounty B-002)): the "King's Book" published days after Charles I's execution. Did Charles write it, or did John Gauden, who claimed it after 1660? | "Milton had a plagiarised prayer (Pamela's prayer, from Sidney's Arcadia) slipped into the book to discredit it." The plagiarism is real; the planting is contested. | Strong: about 60,000 words, far above the length floor; Charles's authenticated letters, Gauden's and Jeremy Taylor's own works as candidates and controls, all public domain | next candidate |
| A-09 | Gowrie conspiracy and the Logan letters, 1600–1608 | "James VI staged the plot to kill the Ruthven brothers." The Logan letters were later shown to be forged by George Sprot | Weak: letters are short; handwriting already studied by Andrew Lang (1902) | parked · short texts |
| A-07 | Casket Letter II (Glasgow letter) | As A-01 | Needs a method that works through translation | parked · method gap |
Sources
- Lasry, G., Biermann, N. & Tomokiyo, S. (2023). "Deciphering Mary Stuart's lost letters from 1578–1584." Cryptologia 47(2): 101–202. doi:10.1080/01611194.2022.2160677 primary_text · open access
- Labanoff, A. (ed.) (1844). Lettres, instructions et mémoires de Marie Stuart, 7 vols. Internet Archive, vol. 1 · HathiTrust primary_text · public domain
- Henderson, T. F. (1889). The Casket Letters and Mary Queen of Scots. Internet Archive primary_text (texts) + interpretive
- Goodall, W. (1754). An Examination of the Letters said to be written by Mary Queen of Scots. vol. 1 · vol. 2 primary_text (texts) + interpretive
- "Casket Letters," Encyclopædia Britannica (1911). Wikisource synthetic
- "The Casket Letters: A Famous Case Reopened." The Historical Journal 28(3). Cambridge interpretive
- NSA, "To Catch a Queen" (declassified 2011). PDF synthetic
- "Thomas Phelippes." Wikipedia synthetic · to be re-anchored to TNA
- National Museums Scotland, "The controversial letters associated with the Mary, Queen of Scots Casket." nms.ac.uk adoption signal
- The Tudor Society, "The Casket Letters" (summarising John Guy, 2004). tudorsociety.com synthetic · to be re-anchored to Guy
- "Casket letters" and "Mary Queen of Scots (Fraser book)." Wikipedia · Wikipedia synthetic
- Northeastern University News (2024), on Estelle Paranque and access to the 2023 letters. news.northeastern.edu synthetic
- Eder, M. (2010). "Does size matter? Authorship attribution, small samples, big problem." Digital Humanities 2010. DH2010 abstract synthetic · method
- "Authorship attribution in translated texts: a stylometric study" (PROPOR 2024). ACL Anthology synthetic · method