972Unsolved Historical Ciphers
m ✕ y b ð a ʓ v ð L ɸ 25 ⸸ † m q ɵ f b v ✕ m ʌ 8 a q
ð w n ð b ð n ✕ q e p ʆ ð a k n r q ✕ ɸ n p b e ɵ e v m v q b s r ð q € 10 f a L o m ð 24 ⁂

Brussels → the Court · homophonic substitution, one glyph per letter · 13 July 1572 · read

A segmenter, not the archive: Mondoucet’s despatch of 13 July 1572

BnF Français 16127 (Harlay 228), Lettres originales de plusieurs ambassadeurs français aux Pays-Bas · Claude de Mondoucet to Charles IX · c. 20 cipher letters with the Court’s decipherments · the one of 13 July 1572 (ff. 60–61) without

Four aligners proved on controls had all sat at the shuffled baseline. The cause was not the manuscript but the front of the pipeline. Read by hand, the cipher is ordinary; the key was built from the volume’s own crib and confirmed against the decipherer’s own gloss.

Daniel Bourdeau · solved

Summary. A Gallica sweep for cipher leaves outside the standard lists turned up BnF fr. 16127, the Court’s file of Claude de Mondoucet’s despatches from the Low Countries, 1571–74: about twenty ciphered letters in one system, most followed by the Court decipherer’s reading, one long letter of 13 July 1572 without. It is on no list — not Tomokiyo’s, Schmeh’s or Dunin’s — and it dates four days before Genlis’s relief column was cut to pieces at Saint-Ghislain and six weeks before St Bartholomew. Four alignment methods, each validated on synthetic controls, had all failed against the readings that do exist and sat exactly at the shuffled-plaintext baseline. The cause was diagnosed here as an artefact of the pipeline’s first stage: an automatic glyph segmenter, the one component never checked against a hand count, split looped d, the crossbarred long-s and the barred o and q into two components each and manufactured a 1.3–1.4 glyph-per-letter ratio, breaking the one-to-one mapping every solver assumed. Transcribed by hand at glyph level, the cipher is an ordinary sixteenth-century homophonic substitution: one glyph per letter, 24-letter early-modern alphabet, a few digraph signs and four nulls. The key was built by hand from the 16 July crib in the same volume — the decipherer’s interlinear glosses on f. 62r and the verbatim reading at f. 64 — and a French 5-gram beam decoder then read the despatch. Status: read. The residue that does not resolve is transcription-limited on a ~450-year-old microfilm scan.

01 The sweep and the document

Gallica’s catalogue was queried for diplomatic manuscripts of 1575–1620 absent from Tomokiyo’s pages, and the hits in range checked against his coverage of the League volumes and the reigns of Henri III and IV. Of the residue, fr. 16127 — original letters of French agents in the Low Countries, 1571–94 — was fetched entire (from the Gallica PDF after the IIIF service throttled) and read on contact sheets. Roughly twenty of Mondoucet’s despatches carry cipher, and almost every ciphered leaf is followed, on the next folio or in the margin, by a contemporary decipherment in the Court’s hand. One stands out: 13 July 1572, ff. 60–61, has no decipherment anywhere in the volume. Its date places it four days before the Huguenot relief column under Genlis was destroyed at Saint-Ghislain outside Mons, and six weeks before the massacre of St Bartholomew. Mondoucet’s office was exactly the covert French contact with William of Orange’s revolt that these weeks turned on; the absence of a Court reading, in a volume where every other ciphered despatch has one, is what made the leaf worth the work.

02 Why four solvers had failed

Three transcriptions had been made — one by hand at glyph level, two by automatic segmentation and k-means clustering of glyph shapes — and four aligners built and validated on synthetic controls of the same size and shape: hard-EM Viterbi, a consistency beam, a soft-EM banded HMM, and the same with word-signs consuming several letters. The controls recovered 44 of 44 and 56 of 58 symbols. Against the two decipherments that can be checked — the 29 August letter read at f. 9, the 16 July block read at f. 64 — every method scored at the shuffled-plaintext baseline.

The clue was a number. On the 16 July pair, which is verbatim where the gloss lets one check it, the automatic segmenter counted 1.3–1.4 connected glyph components per plaintext letter. That was read at the time as the cipher doing something — heavy nulls, or letters written with more than one separable stroke — and it was what defeated the aligners, since a solver that assumes one glyph maps to one letter cannot align a stream with 35 % too many symbols. But the ratio was measured on the segmenter’s output, and the segmenter was the one component never validated against a hand count. Hand transcription of the glossed words settles it: under the f. 62r gloss leur délivrance, the number of pen-lifted glyphs equals the number of letters, and the same holds under pouvez and au 30 and across the 16 July block wherever the f. 64 reading pins the letters. The excess was an artefact — the segmenter split looped d, the long-s with crossbar, and the barred o and q into two components each, and attached dots and bars as separate glyphs. One glyph is one letter. The archive was consistent all along; the pipeline’s first stage was not.

03 Building the key by hand from the 16 July crib

The volume contains its own crib. The 16 July 1572 pair (ff. 62–64) is enciphered in the same system and glossed: on f. 62r the Court decipherer wrote several plaintext words above the cipher (pouvez, leur délivrance, la bruslation, nous maintenant, au 30), and f. 64 (“16 juillet 1572, déchiffré de la précédente”) gives the whole block in clear, 17 lines and 863 letters. Anchoring the glossed words fixes about twenty letters at once. Carried into an anchored HMM alignment of the hand transcription — forward–backward and Viterbi, banded around the gloss anchors, with a per-glyph null rate — they give a table of homophone counts per glyph and a small null set { * , C , mq , j }. The cipher is homophonic with a 24-letter early-modern alphabet (i=j, u=v), a few digraph signs (the sign transcribed 23 = qu, crossed-phi = ll), and those four nulls. Reading is then a decoding problem, not a search: a beam decoder (width 300) combines, per glyph, the probability of each letter from the aligned counts with the transition probability from a 5-gram model of sixteenth-century French, and emits a null marker where even the best letter is improbable.

The decisive control is internal. Decoding the 16 July block with the recovered key — whose clear text is independently known from f. 64 — reproduces the decipherer’s own gloss. The line under which f. 62r reads leur délivrance decodes to

“… leur delivrance des espagnols a ce moyen …”

and the surrounding lines read continuously with it (“… penser que le roi lui a prester ses affaires …”, “… envoyer afin … clement adverti du premier … de son armée assemblée de laquelle aura …”, “… prelate teste en flandres …”). The gloss match is not a coincidence a wrong key could produce; the key carries.

04 The 13 July despatch

The despatch is only partly enciphered. Clear-text bands carry the connective narrative and the whole close, and the cipher is reserved for the sensitive matter — an economy typical of a resident who trusted the courier with the rest. The clear bands read directly. On f. 60v, a Spanish informant’s report:

J’ay esté asseuré par Espaignol mesme qui cy vient qu’il n’y a pas troys mil hommes de pied & cinq cens chevaux cy tout. Toutesfois ilz demourent encores là arrestez. Et dit on que du costé de Mastricq se marche quelques regymens d’allemans nouvellement leuez…

f. 61r closes in clear: Sire je supplie le Createur qu’il vous donne … très longue vie. De Bruxelles ce [Ve] jour de Juillet 1572, signed de Mondoucet, over vostre treshumble tresobeyssant et tresaffectionné serviteur et subiect.

The ciphered lines — 22 on f. 60r, one on f. 60v, six on f. 61r — decode in stretches with the same key. They speak of religieuse, Espaigne, the saint père and the pape, Escosse, Angleterre, troupes de terre, catholique, ministres, liberté, garnison; f. 61r gives … le prince … service du roi, … parenté entre les autres cour[s], … avaient esté capitulé … réduit. The residue that does not resolve is transcription-limited — individual glyph reads on a ~450-year-old microfilm scan, not a failure of the key.

05 The lesson

A ciphered despatch that four validated aligners could not touch was an ordinary homophonic substitution all along; what defeated the solvers was an unvalidated segmenter that manufactured a 1.3–1.4 glyph-per-letter ratio and broke the one-to-one mapping. The lesson is narrow and worth stating: never trust a glyph-per-letter ratio taken from an automatic segmenter you have not checked against a hand count on a passage of known plaintext. What remains is a clean retranscription of the ciphered lines to a fixed glyph accuracy and the collation of the clear bands against the Didier (1891) edition of Mondoucet’s correspondence.

06 Sources