Candidate key for ASV Segr. Stato Spagna 1A, ff. 70r–73v

← The Farnese letter · Edited Italian text

The candidate key reads 7 3 4 5 7 4 9 0 3 5 as dopo la: 73=d, 4=o, 57=p, 4=o, 9=null, 03=l, 5=a. This starts the first enciphered passage of ASV_i1025_SdS_Spain_IA-2, the target of The Vatican Challenge Part 5.

The key was recovered from the public transcription. The reading provisionally identifies the letter as Cardinal Alessandro Farnese to Giovanni Poggio, the papal nuncio at Charles V's court, Rome, April 1542. The precise day remains uncertain. Manuscript photographs supplied through DECODE record 92 were examined on 15 September 2026. The edition uses the corrected transcription.

The candidate is a monoalphabetic variable-length substitution: each of the 18 letters in the recovered alphabet has one code, of one or two digits. Dotted pairs encode additional syllables, words, and titles. The candidate letter alphabet has no homophones (alternative codes for the same letter).

Code boundaries must be inferred from context. For example, 5 7 can be p or a n, 8 0 can be t or e s, 0 3 can be l or s m, and 7 5 can be g or n a. The tables below give the proposed assignments.

Files and notation

The edition file guide describes the edited text, alignments, and automatic readings. The transcription appendix contains the source texts and manuscript decisions. The results supplement contains the recovery results, controls, and blind rereads.

In key.txt, ce and ci stand for che and chi. The title symbols are # for S.S.tà (the Pope), & for S.M.tà (the Emperor), and % for V.S. (Poggio). _ is the null, which emits no letter. A ^ follows the dotted digit: 27^ is 2 7̇, and 2^7 is 2̇ 7. These are notation conventions, not additional cipher assignments.

Regular codes

Plaintext Code Plaintext Code
a 5 l 03
e 8 t 80
o 4 d 73
i 6 c 27
u / v 1 p 57
n 7 b 72
r 2 f 37
s 0 g 75
m 3 z 07
et 81 null 9
  • 9 represents no plaintext character: it is a null. In the proposed reading, it occurs between words (dopo·la partita del Montepulciano·il·Marchese) and within words (tu·te = tutte, cir·ca). Here · marks the null. It can also prevent ambiguous grouping: 8 9 0 separates e and s, whereas 8 0 can be t (e·sendo, e·ser, pote·sero, vole·se, e·stata). Likewise, in ca·desse, 5 9 73 separates a from d, preventing the first two digits from forming 57 (p).
  • The proposed reading generally omits double consonants (dopo, tuto, posibile, sumario) and h (auere, a = ha, omini). The candidate key has no separate codes for h or q; qua, que, qui, chi, and che have dotted codes below. For example, quando = 2̇7 7 73 4. u and v share 1.
  • Read as separate single-digit letters, the two-digit codes give 03 = sm, 73 = nm, 80 = es, 57 = an, 27 = rn, 72 = nr, 37 = mn, 75 = na, 07 = sn, and 81 = eu. Some sequences favor a two-digit code, but letters can also cross word boundaries. Code boundaries must be inferred from context. For example, 7 5 1 6 is read as navi (n a u i) in le cento navi, rather than g u i.
  • Some abbreviations are spelled with letter codes: Mons. (3 4 7 0), Reu. (Rev.mo), Apostol., and the Ces. in S.M.tà Ces. (the S.M.tà code followed by 27 8 0).

Dotted codes

The public transcription uses d^. for a dot above digit d. The key search treated ^, marks as dots too. The error-tolerant decoder also considered whether to retain or ignore ^- marks. These conventions do not establish which marks are dots on the manuscript.

In the candidate key, the position of the dot within a pair matters. For the pairs 27, 72, 57, a dot on the first digit gives qua, que, qui; a dot on the second gives che, chi, non. In plain-text notation, 2^7 means 2̇ 7, while 27^ means 2 7̇.

The examples below are edited contextual readings. The title codes are identified by their referents; the reading uses S.S.tà for the Pope.

Code Proposed meaning Contextual examples
2 7̇ che il Mar·che·se d'Agilar, tutto quello che patiranno le sue forze, secondo che V.S. scrive
2̇ 7 qua del quale, il quale, qual parte, li quali, quando, qualunche
7̇ 2 que con questa, in questo caso, queste considerationi, quello, quelle parti, adunque
7 2̇ chi che richiedano, carichi necessarii, altri luochi maritimi, da chi ne desse loro
5̇ 7 qui in obedienza et qui·ete lo stato suo, essendo stato ese·qui·to dal vescovo di Modena, firmato qui in Roma con Mons. di Granvela
5 7̇ non non si vede, di non voler lasciar indrieto cosa alcuna, quando non fussero intratenuti, non meno per amor suo privato
0̇ 5 S.S.tà / N.S. (the Pope) il Marchese d'Agilar è stato con S.S.tà et li ha conferito il parer del Principe Doria, la risposta di S.S.tà, S.S.tà se ne riposa … nella prudenza della M.tà Sua, S.S.tà parata di concorrer
4 0̇ S.M.tà (the Emperor) in verso di S.M.tà, nella prudenza della M.tà Sua, ad instanza di S.M.tà Ces.
6̇ 0 V.S. (Poggio) secondo che V.S. scrive per le sue, come V.S. sa, la risposta che V.S. vedrà per lo incluso capitulo, et V.S. ne sarà avisata

In these examples, · brackets a syllable to show its position within a word.

Unresolved marks and proposed corrections

The following marked signs remain unresolved in the edited reading:

Mark or sequence Occurrences Folio and context
9̇ 2 072r: per dar 9̇ 4̇; 072v: condotta da 9̇ 1 0
0 1̇ 2 072r: at the opening and before ne ho ricevuto
4̇ 1 072r: per dar 9̇ 4̇

The two marks in 9̇ 4̇ are counted separately. 0 1̇ stands both times at a clause boundary (et datone noticia 0 1̇ perché così era ricercata S.S.tà; instanza di S.M.tà 0 1̇ ne ho ricevuto ultimamente la risposta) and may be a word or title code for the place or person the notice went to and the answer came from (a/da Spira, the Diet; Modena, Morone); the dotted null might identify the power that hired Gonzaga. These interpretations remain open. The manuscript check resolved the 7̇ 3 in da che parte (072r): the page has 2 for a later 3 in the public transcription, so the stretch reads da che parte with one stray dot.

Other marks have proposed explanations as stray or displaced dots, for example 6 0̇ read as 6̇ 0 in V.S. ne stia sicura, 2̇ 7̇ read as che, and 4 0̇ read as o[l]tre rather than S.M.tà in le quali sue forze però o[l]tre che per lor stesse non sono grandi (071v, after the blind reread). The dot decisions cover 45 places where the edition ignored or supplied a dot. The check found nine dots missing from the transcription: in quele, il bisogno S.S.tà, insieme non, modi che, le quali, di S.M.tà Ces., di S.S.tà di conservar, calumnie che, and il qual luogo. Sixteen transcribed marks are not dots on the page. Seven real dots stand where the reading needs none, such as a stray dot within Signor or the misplaced dot in 2̇ 7 for Marchese. The alignment still records 22 mark operations at the clerk's dots, blots, and unresolved places.

The manuscript check produced 129 entries in the correction ledger; two remain open and are not applied to the corrected transcription. Five groups written once on the page were transcribed twice, adding 21 digits. Other corrections include fourteen omitted digits, tokens without a glyph, two swapped pairs, digit confusions, and above-marks. The confusions include 1/2, 2/3/7, 0/8/9, and 0/6. Some apparent dots belong to a neighboring digit or the line above.

The review covered all 196 ranked classifier suspects and all flagged differences in digit count. The decision ledger records the individual judgments. The review did not check every glyph the classifier accepted. Some remaining departures may be the clerk's slips or blots: a 2-form stands where n is needed in Go[n]zaga, and 3 7 = f where proprii needs 5 7 = p. Other readings remain conjectural.

alignment.txt records how the edited decipherment accounts for the 6,562 corrected tokens under the key. It finds a least-cost alignment of an already supplied text. Bracketed supplements receive a reduced cost, and gaps can absorb digits without cost. The alignment makes departures visible but does not independently establish that a reading is correct.

The recorded operation counts are 59 supplied letters, 45 digit substitutions, 37 unreadable digits filled, and four two-digit codes with one digit missing. They also include 22 mark operations (16 ignored, six supplied), six dropped digits, ten tokens read as nulls, and seven entries labeled as alternative choices. One label is wrong: on 71r, 3/7>1 supplies the unbracketed u in sua parte. Neither offered alternative is 1, so this is a substitution, not an alternative choice. The published tally therefore understates substitutions by one and overstates alternative choices by one.

Square brackets mark letter departures in the edited text, with that known exception. The alignment also records changes that brackets cannot show, such as ignored dots and dropped digits. Of its 45 labeled substitutions, 37 use the listed confusion classes. The other eight (1/3, 1/8, 1/0, 5/6 twice, 6/5 twice, 6/1) need particular caution. The separately mislabeled 3/7>1 also falls outside those classes.

The alignment reads 5,993 tokens exactly and 365 as the null 9, consumes 170 tokens in departure categories, and leaves 34 in gaps. These are token counts, not operation counts. Both the edition and transcription changed after the manuscript check. The earlier alignment therefore describes a different text, not the current edition against the old digits. See the article's comparison.

The automatic reading in reading.txt records 165 edited units containing 170 operations:

  • 78 digit substitutions, including six selections of an alternative already offered by the transcriber;
  • 28 above-marks ignored and 14 dots supplied;
  • four digits inserted and two dropped;
  • 44 fills of unreadable ? tokens.

Excluding the 44 fills leaves 126 operations over 6,577 source tokens, or 1.9%. This is an edit-operation ratio, not a measured transcription error rate. Multiple operations can affect one unit.

The automatic reading retains rejected proposals, including santa where the edited reading selects spesa, and g u i where it selects navi. It also inserts title codes into ordinary words. The edited plaintext therefore differs from the automatic reading; reading.txt records the decoder's proposals, and alignment.txt records the editorial reading's.

Relation to Meister's keys

The candidate alphabet differs from Meister 176/2, Cifra ultima con Mons. Poggio mandata per il Montepulciano (dated 1538–42 by Meister). For example, that key gives a=3; the candidate gives a=5.

The heading connects that key's delivery with Giovanni Ricci da Montepulciano, but does not establish which journey carried it. The edition's cleartext XXV is an insecure shortening: further numeral strokes may give xxviij (28), consistent with the published date of 28 March 1542. See the article's note on Ricci. The different alphabets establish neither the delivery date nor the order of use.

Structural parallels appear in the Argenti Cifre vecchie usate, nos. 7, 13, and 14 (Vitelli, Colonna, Marignano), printed in Meister's 1906 study, pp. 179–182. These are comparisons, not a matching historical key. The instruction "si ha da intendere per discretione" belongs to another key (no. 9, p. 180) and illustrates reliance on context.

Recovery and checks

The recovery was LLM-assisted. A search over candidate keys and code boundaries used a character five-gram model of sixteenth-century Italian. Four of six restarts in the first successful search found the same 18-letter key. A tighter search, with one code per letter, one null, and no word units, recovered it in two of three restarts. The dotted codes and et were inferred from their decoded contexts. The results supplement gives the full search and control results.

Separate searches on 070r–071r and 071v–073r recovered the same letter key and null in all twelve restarts. Clean and noisy synthetic ciphers each yielded their known letter key and null in all three restarts. Their Italian partly overlapped the training text, so they are not independent held-out tests. These controls did not recover the dotted codes or et. The shuffled ciphertext scored −1.351 per retained token, compared with −0.783 for the real-text key. These scores rank readings under the shared language model, not probabilities that a reading is correct. Deterministic reruns reproduced the recorded keys and scores.

Two blind rereads helped revise the edition. Both readers were fresh GPT-6 Astra runs, given the key and a language-model-selected strict segmentation of one folio. Their packets used the first listed alternatives and omitted unreadable tokens and short fragments. They did not contain the edition or manuscript images. Their independence is therefore limited to the subsequent reading task. The manuscript check supplied separate evidence about the ink, while the remaining conjectures and unresolved codes still need confirmation.