Raince to Montmorency (1526)

Summary. Nicolas Raince was the French embassy’s secretary at Rome, and BnF fr. 2984 (Béthune 8509) keeps eight of his despatches of 1526 in cipher. Tomokiyo published the key for them in 2020, DECODE holds two of the letters as decrypted, Mignet noted in 1886 that the one letter “entirely in cipher” has its clear duplicate eight leaves earlier in the same volume, and Bourrilly printed the ciphered passages of two more in 1901. What has never been read anywhere found is most of the 13 May 1526 letter to Montmorency and the lower part of the 20 November one: about 106 lines. The obstacle was never cryptanalytic. It was that the published key had been read off Tomokiyo’s table by eye and slipped one column, so that l, m and n were each a place wrong and every reading built on it was corrupt. Measured off the image instead, the table resolves a control line glyph for glyph, and 86 of the 106 lines have now been read by hand from the microfilm. The 13 May letter reports the Emperor’s side of the negotiation nine days before Cognac — someone qui estoit en la court de Savoye est party pour venir icy, the Castilians and the Bourguignons, a plan to prendre le chemin de Valence, a capitulation for which ilz seront cause de la destruction, the plague in Rome, and a named man taken un jour, en plaine place pres du palais. The 20 November letter turns on a pontificate that seroit la cause de la totale ruine de sa maison. Status: read in part. 86 of 106 lines read glyph by glyph, with gaps left as gaps; 20 lines carry only a machine decode and are excluded from the edited text. One glyph, occurring ten times and always at a clause boundary, is in no published table.

01 The documents

BnF fr. 2984 is volume 8509 of the Béthune collection, the Court’s file of correspondence from Rome in 1526. Gallica has it as a black-and-white microfilm (ark btv1b90598430, 85 views, every canvas labelled “NP”). Eight items are in cipher:

no.	page	date 1526	to	cipher	read where
6	25	1 Aug	Montmorency	last 7 lines	contemporary interlinear decipherment on the leaf
7	29–31	13 May	Montmorency	p. 29 whole, p. 30 whole, p. 31 upper half; clear close	gloss on the first ten lines of p. 29 only; the rest read here
8	33–37	9 June	the King	the whole letter	clear duplicate at p. 17 ff. (Mignet 1886)
10	41–46	17 June	the King	most of the letter	interlinear decipherment; cipher parts printed by Bourrilly 1901
11	47–50	20 Aug	Montmorency	parts	cipher parts printed by Bourrilly 1901
24	105	20 Nov	Montmorency	40 lines after 4 in clear	gloss on the first dozen lines; the rest read here
28	121–123	7 Dec	Montmorency	c. 18 lines	dense interlinear decipherment on the leaf
fr. 3091 no. 11, 12 Aug 1529, is DECODE 3696, decrypted, in Raince’s second key.

Three lines of cipher from BnF fr. 2984 p. 30, small graphic signs written continuously — Fr. 2984 p. 30, the first three lines: *…ction, et à vous faire tous services et plaisirs · qu’il est. L’empereur est esbay de ce que … n’a escript autrement, aucune part deça, depuis son retour…* The bold double-barred sign near the middle of line 1 is the glyph that is in no published table (Gallica, btv1b90598430, view f17, left-hand page).

02 The key had been read off by one column

Tomokiyo set out the 1526 key in French Ciphers during the Reign of Francis I (2020) as an image: a header row of plaintext letters over up to four rows of cipher glyphs, the columns about 78 px apart. It is a monoalphabetic substitution with homophones, three nulls and three word signs. Reading which glyph sits under which letter by eye is the whole difficulty, and the first pass here slipped a column.

Doing it by measurement instead settles it. Locating the white header letters in the magenta band, then labelling the ink blobs below and assigning each to the nearest header column, every blob falls within 15 px of a column centre — a fifth of the column spacing. The measured table is:

Contact sheet of every glyph cut out of Tomokiyo's key table and labelled with its plaintext letter — Every glyph cut out of the key image and labelled by the column it measures into. Three homophones for a and *u/v*, four for e, three for i, two each for l, o, p, r and s; no glyph at all for x; three nulls.

The correction. l is the 9, not the round s; m is the round s, not the long f; n is the long f. The first pass had these one place to the left. H measures under s, not under r where it had been put.
Why it matters. That single shift is what produced dongUeaent for longuement, egsembde for ensemble and SonseigEur for monseigneur in the first machine draft. Every file derived from the shifted table was discarded.
Confusable pairs. p has a T and a circle with a dash to its right; r has a 4 and a dash then a circle. The two are near mirror images and are the easiest pair on the leaf to confuse. The most frequent single glyph, a circle at the foot of a barred stem, is an e.
Word signs. A looped L with a tail = con; two small circles joined by a curve underneath = l’empereur; a cursive y with a descender = le pape. The letter y is E, so a y-shaped glyph is the word sign, not the letter.

The control. With the measured table, p. 29 line 20 resolves glyph for glyph and leaves nothing over:

∧ λ n v ✕✕ 7 ∧ ℓ ꝛ I ℓ 7 f 9 o ⊡ ꝛ v 4 ℓ ϝ 7 ∧ o R ꝛ E ψ ny 7 ∧ ℓ T o 4 ℓ E T ꝛ v 4 R 7 f ✕✕ 4 I ⊡ E
→ s [null] qui estoit en la court de Savoye [null] est party pour venir icy

The line uses a, c, d, three of the four e homophones, two of the three i, l, n, o, p, q, r, s, t, u/v, y, and two of the three nulls. That is the evidence the table is right.

03 pp. 29–31, not ff. 29r, 30v, 31r

The numbers written on successive Gallica views run 29, 31, 33, 35, 37 — two apart, not one. Read as folio numbers that would mean half the volume’s openings were never filmed, and every letter in it half missing. They are page numbers, written on odd pages only: view n shows pages 2n−4 and 2n−3, 85 views cover about 167 pages, and the digitisation is complete. So the 13 May letter is three consecutive pages, 29, 30 and 31, and the 20 November one is p. 105.

The text settles it independently. Page 29 ends … et n’est possible … monstrer plus a[ffe]- and page 30 opens -ction. Both halves of that join were read glyph by glyph before it was noticed.

04 Reading a microfilm

The film is legible at about 2× but the ink is grey and the hand is small, so the difficulty is not the cipher but the transcription. Three things made the reading possible.

One image per manuscript line, cut into four overlapping quarters and stacked, so a whole line is read in one view without losing magnification.
Numbered tokens. Connected-component segmentation gives 5,629 glyph tokens; drawing each token’s box on the line and numbering it means a hand reading is written as one character per token and lands aligned to the segmentation with no guesswork. Every line was checked against its token count before it was kept.
A scaffold, not a reading. A classifier over the token images, combined with a space-free period-French 6-gram beam, proposes a letter for every token. It is right on about 85–90 % of them, which is enough to speed a hand reading and nowhere near enough to be one.

One manuscript line at 2x with every glyph boxed in red and numbered, in four stacked strips — p. 30 line 8 with every segmented token boxed and numbered. The hand reading of this line is *[null] respondu à ce que ce qu’on a envoyé d’icy et de V[null]enise par deu-*, running on into line 9 as *-ers ledict seigneur*. Token 39, a K, is a null standing inside the word *Venise*.

Where the ceiling is, measured. The hand readings double as training data, so the classifier can be scored honestly by leaving one hand-read line out at a time:

training data	leave-one-line-out per-token accuracy
the map’s own confident tokens only	0.840
+ hand labels, weight 1	0.856
+ hand labels, weight 8	0.901 (peak, at 32 hand-read lines)
+ hand labels, weight 30	0.860

That is the whole argument for doing it by hand. At 0.90 per token a six-letter word survives at 0.53, which is why only about a third of the machine text segments into real French words however the decoder is tuned. The residual confusions are between glyphs that really are similar at this resolution — ψ (e) against 4 (r), the cursive ℓ (t) against 9 (l) against the lollipop (d), the caret (s) against the vee (u) against × (a) — not between glyphs the key leaves ambiguous. More labels will not fix that; better images would.

05 The letter of 13 May 1526

Nine days before the League of Cognac was signed. The clear close on p. 31, transcribed here for the first time, dates and signs it — De Rome, ce xiij jour de may, 15c xxvj … Nicolas Rainse — and reports that Rome expects Andrea Doria at Civitavecchia that day or the next, with six galleys and two brigantines. The three ciphered pages that precede it read, in part:

… ne cessera jamais qu’il n’ayt destruict ledict seigneur … — and, a line later, d’autres propo[s] … ains de vert[u] et mensonges.
… ont receu … de Castille, qu’ilz disent …; ensemble … les bourguignons; and, of a great lord, les castilliens li[sent] au (con)traire que lesdessusdict[z] sont …
… delib-eré prendre le chemin de Valence … à ceste fin, et qu’il ne voul-oyent qu’il feust … si grant seigneur.
… et qu’il[z] seront cause de la destruction, de la part d’avoir faic-t capituler avec … — the likeliest referent being the Treaty of Madrid of January 1526, though the letter does not name it.
Ledict He[rr]er[a](?) doit partir de dans deux ou troys jours pour s’en retourner — the name is not settled —; and …s qui estoit en la court de Savoye est party pour venir icy.
p. 30: (L’empereur) est esbay de ce que … n’a escript autrement, aucune part deça, depuis son retour … en son royaume; et m’en a parlé plusieurs fois.
… que mondict seigneur l’Admiral luy av-oit dict que seroit celuy qui y tiendroit. J’ay tousiours respondu à ce que ce qu’on a envoyé d’icy et de Venise par devers ledict seigneur …
… deux ou trois fois qui a[voit] deu venir à Rome … disant prendre tout pour le mieul[x] … Bien est vray, monsei-gneur…
p. 31: … la peste qui est au logis …; …s foys escript des advis qui viennent aux imperiaulx par leu[rs] …; … es grant dommaige en toutes les sortes qu’on pourroit …; [il] y ayt grant obligation pour les dommaiges qu’ilz disent ledict avoir…
and, near the end, a named man: …t un jour, en plaine place pres du palais, un nommé Pietro Anthonio … secretaire ou serviteur dudict, qui estoit icy pour luy … followed by propos … quel estoit cause de tout le mal … deux ou trois fois.

06 The letter of 20 November 1526

Two months after the Colonna raid on the Vatican, and five months before the Sack. The ciphered part turns on the Pope:

… qui a … peust … faire s[e]cre[te]ment · et … pontif-icat seroit la cause de la totale(?) ruine de sa maison, de-…

With, in the same passage, …me, et s’en fuire; et que encores … luy feist … la grace; … et que … l’on penseroit … y respondre de part de nature; …t jamais pensé que les choses … prinst, et le for[t]-; sincerement … qu’il estoit … en besoing; … et que l’i[n]convenient … peril à qu[oy] …; and et aussi qu’il estoit … bien cler que n’estoit … dicte de. “Sa maison” is the Medici, and the pontificate is Clement VII’s.

07 What remains open

Twenty lines of the 106 carry only the machine decode. They are kept, marked, in trans/reading.txt and left out of the edited text rather than dressed up as a reading.
Gaps inside the 86. The edited text leaves … wherever the glyphs did not resolve. Nothing is supplied by conjecture; a wrong confident reading would be worse than a hole.
One glyph is in no published table: a bold vertical stem crossed by two horizontal bars, the lower one longer and projecting right. It occurs ten times across all four pages and every occurrence falls at a clause boundary — between plaisirs and qu’il, between lectres and qui est, opening a sentence. It is not the barred-b g, not the barred-stem e, and not any of the three nulls. Most likely a point or a fourth null; it is written · in the edited text.
The p. 29 / p. 30 join reads …monstrer plus a[ffe]- + -ctio?e, where the glyph after ctio measures as a plain triangle (f) rather than the n the sense wants. Left with a query.
Two personal names are not settled: messire Ca?in, who appears twice on p. 30, and the agent read He[rr]er[a] on p. 29.
What would finish it is not modelling but light. The Gallica item is a black-and-white microfilm; reader’s photographs of pp. 29–31 and 105 would let the segmentation run clean, and the measured ceiling above says that is worth more than any further training data.

08 Method

Key measured, not read. Header letters located in the key image, ink blobs labelled and assigned to the nearest column, every assignment within 15 px of a column centre. Contact sheet of every labelled glyph kept.
Verified on the leaf against a control line that resolves with nothing left over, using sixteen of the alphabet’s letters and two of its three nulls.
Segmentation by connected components on a background-normalised binarisation, tokens assigned to lines by a y-histogram, x-overlapping components merged; 5,629 tokens over the four pages.
Cluster-to-letter map hill-climbed against a space-free period-French 6-gram, with each cluster restricted to the letters its glyph shape allows. Restricting the moves is what stops the collapse into a degenerate map that defeats an unconstrained solver on a stream with this much segmentation noise.
Hand reading line by line against numbered-token images, one character per token, length-checked against the segmentation.
Honest scoring by leave-one-line-out over the hand-read lines, since a model trained on its own output otherwise scores itself.

Everything is in the repository: the measured key and reading method in raince/CALIBRATION.md, the account in raince/NOTES.md, the hand labels in raince/gold/labels.txt, the edited reading in raince/trans/reading_edited.txt, the line-by-line text with each line marked hand or machine in raince/trans/reading.txt, and the clear close of 13 May in raince/trans/p31_clear_close.txt.

09 Sources

BnF Français 2984 (Béthune 8509), Gallica ark btv1b90598430, microfilm, 85 views.
S. Tomokiyo, French Ciphers during the Reign of Francis I, cryptiana.web.fc2.com, first posted 2 November 2020: the 1526 key (francisRaince.png) and the 1529 key, crediting Desenclos (HistoCrypt 2018) with locating the letters. He gives no plaintext.
DECODE records 4228 (fr. 3040 f. 21, 1526) and 3696 (fr. 3091 f. 19, 1529), both decrypted. Nothing from fr. 2984 is in DECODE.
F.-A. Mignet, Rivalité de François I^er et de Charles-Quint, 2nd ed. 1886, t. II p. 211 n. 1: the 9 June letter “l’original en chiffres f. 33 … le déchiffrement f. 17”. His “f.” are the same numbers written on the leaves, which section 03 shows to be pages.
V.-L. Bourrilly, “La première défection de Clément VII à la ligue de Cognac”, Bulletin italien 1 (1901) pp. 213–229, appendix II: the ciphered passages of the 17 June and 20 August letters, which he read himself.

Nicolas Raince to Montmorency, Rome, 13 May and 20 November 1526