Mabeyn to Yemen, system B: 21 telegrams in two-digit numerals with the clerk's Arabic letters over them, 1896(?)-1900 (HPBK, Hilmi Pasa collection)

TASK: help read Unsolved Catalogue no. 2646 of the cyphersolver project
This page: https://cyphersolver.com/solve/2646.html

You are helping a public research project read historical cipher letters that nobody has read yet.

THE CIPHER
  Title:          Mabeyn to Yemen, system B: 21 telegrams in two-digit numerals with the clerk's Arabic letters over them, 1896(?)-1900 (HPBK, Hilmi Pasa collection)
  Date:           Jun 1898 - Nov 1900 (580 and 587 are catalogued 8 Aug 1896; 583 and 585 undated)
  Correspondents: Mabeyn-i Humayun cipher clerks (Mehmed Kamil named on 518 and 583) -> Yemen Province (Huseyin Hilmi Pasa)
  Place:          Istanbul (Yıldız) -> Sanaa
  Language:       Ottoman Turkish plaintext (Arabic script) under a two-digit-numeral cipher; telegram forms (expected)
  Held at:        Koc University (Istanbul), Hilmi Pasa Belgeleri Koleksiyonu (CONTENTdm HPBK), items 495, 498, 518, 524, 527, 536, 541, 546, 562, 575, 577, 580, 583, 585, 587, 602, 623, 644, 652, 675, 1761; 21 sheets, one telegram each. Dates as catalogued: 675 (20 Jun 1898), 644 (28 Jun), 652 (3 Jul), 623 (4 Jul), 602 (12 Jul), 498 (16 Aug), 541 (7 Sep), 495 (23 Sep), 562 (6 Feb 1899), 527 (24 Apr), 577 (30 Apr), 546 (4 May), 524 (20 Aug), 536 (21 Nov), 575 (6 Feb 1900), 518 (14 Nov 1900), 1761 (18 Jun 1898), 580 and 587 (8 Aug 1896 as titled; 587's date field says 1898), 583 and 585 undated. Not in this entry (other entries): 533 and 688 (2575, same system B), 483, 576, 578, 599 (2577). Clear sheets of the same days that serve as cribs: 725 and 748 (28 Apr 1901) and 726 and 764 (5 Jun 1901) against 578/599 and 576 (entry 2577); 1070 (6 Feb 1900) against 575 (this entry); the others of the 36 items of old lead 2578 are plain letters.
  Images/records:
    - HPBK item 495: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/495
    - HPBK item 498: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/498
    - HPBK item 518: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/518
    - HPBK item 524: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/524
    - HPBK item 527: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/527
    - HPBK item 536: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/536
    - HPBK item 541: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/541
    - HPBK item 546: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/546
    - HPBK item 562: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/562
    - HPBK item 575: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/575
    - HPBK item 577: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/577
    - HPBK item 580: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/580
    - HPBK item 583: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/583
    - HPBK item 585: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/585
    - HPBK item 587: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/587
    - HPBK item 602: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/602
    - HPBK item 623: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/623
    - HPBK item 644: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/644
    - HPBK item 652: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/652
    - HPBK item 675: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/675
    - HPBK item 1761: https://cdm21054.contentdm.oclc.org/digital/collection/HPBK/id/1761

WHAT IS KNOWN
  Status: System B of the Mabeyn-Yemen telegrams. Seen on the page images (14 Nov 2026): two-digit numerals with one Arabic letter over most of them on 17 sheets; 587 and 602 numerals only. Catalogued as not deciphered; no reading, key or edition found. A partial reading (letters) is on the sheets; crib sheet for 575 is 1070.
  Outcome so far: not yet attempted by the project
  Prior-art check: open as far as checked (14 Nov 2026): known.py (HPBK, Hilmi Paşa, Mabeyn, cdm21054) and web search found no study of these Mabeyn-Yemen telegrams; Tomokiyo, DECODE and HistoCrypt not searched beyond known.py. The Cryptologia 2023 and Karabacak Yılmaz papers cover Ottoman diplomatic ciphers 1797-1877 from BOA, not this palace-to-province traffic
  Check first: Transcribe 541 (grid, 8 columns) and 585; tabulate numeral to letter; test the mapping on the other 17 sheets and on the six sheets of 2577; read the letters as Turkish (Arabic script) and compare 575 with the clear 1070.
  Class A: a sibling letter with a decipherment is in the same volume or series, so matching it is the first move.

STEPS
0. Set up. If you can run a shell, fork and clone the repository and work inside it:
     git clone https://github.com/dbourdeau/cyphersolver.git
   Then read, in this order: CONTRIBUTING.md (how a contribution is sent and credited); README.md, sections
   "Repository layout" and "Conventions" (how readings are measured and named); and
   lang/README.md (shared language models).
   Put all your work in targets/<writer><year>/ (a new folder, e.g. sega1593): transcription, key, decipherment, scripts. Write a NOTES.md that says
   what the document is, the system, the key, the reading, and what is open and why. Keep WORKLOG.md there too:
   one dated line per step (transcription, hypothesis, solver run, crib, key found, control), failures included.
   Do not edit docs/ (the site is generated and rebuilt on merge) and do not commit large images: link them.
   If you cannot run a shell, do the same work in files you keep and send them with the issue below.
1. Make sure it is still unread. Search the web and the literature for the shelfmark, the correspondents and the date:
   printed editions and calendars of the correspondence, Satoshi Tomokiyo's cryptiana pages
   (https://cryptiana.web.fc2.com/), the DECODE record and the files attached to it, and the project notes above.
   If a decipherment already exists, stop and report where it is. That is a useful result.
2. Get the images. For Gallica items, a page image is
   https://gallica.bnf.fr/iiif/ark:/12148/<ark>/f<page>/full/full/0/native.jpg (page numbers from the manifest).
   DECODE records (de-crypt.org) need a free account: if you cannot log in, ask your human to download the images for you.
3. Look for key material before attacking the ciphertext, in this order: a decipherment written between the lines or
   in the margin; a clear copy or a deciphered sibling letter in the same volume or DECODE series; a key sheet from the
   same office and years (DECODE key records, Tomokiyo's articles, printed editions); clear passages in the letter
   itself that can serve as a crib.
4. Transcribe exactly what is on the page: one token per sign or number group, line by line, with folio and line
   numbers, unclear signs marked with ?. Keep the transcription as a file. It is worth sending even if nothing else works.
5. If no key exists, identify the system (count the distinct symbols, tokens and repeats: simple or homophonic
   substitution, nomenclator with code words, syllabary, two-digit figures) and recover the key from the ciphertext
   with a solver scored on a language model of the period (Ottoman Turkish plaintext (Arabic script) under a two-digit-numeral cipher; telegram forms). In the clone, use the shared models in lang/
   (from lang import lm; lm.load('<model>'), lm.best_language(text)) rather than building your own; add a corpus there
   if none fits. Before trusting a result, run the same solver on a control text you encipher yourself in the same
   system and at the same length. If it cannot solve the control, its output on the real letter means nothing.
6. Measure the result against the project's bar (README.md, Conventions). A text is "complete" only when (1) at least
   95% of its cipher tokens give sense, counted from a reading file, not estimated; (2) no gap is left untried;
   (3) every letter in the item is covered; (4) what stays open is scattered code groups or names, or pieces blocked
   from outside (no key material, too short, illegible, needs the archive). Otherwise it is "partial": say which
   pieces are open and why. Mark every uncertain word with [?] and leave unread groups as their numbers or signs.
   Never fill gaps with plausible guesses. A failed attempt that records what was ruled out is still a contribution.
7. Name how the text was obtained with one of the project's methods (do not write just "solved" or "read"):
     - ciphertext-only break: The key was rebuilt from the ciphertext alone: frequencies, solvers, pattern words, the order of the key. No plaintext and no existing key was used. If someone had solved the same text earlier and it was found only afterwards, the write-up says so: an independent re-solution.
     - key from external plaintext: The key was rebuilt by matching the ciphertext against a plaintext found elsewhere: a printed edition, a calendar of state papers, a copy of the same letter in another archive.
     - key from adjacent plaintext: The key was rebuilt from plaintext next to the ciphertext: a partial decipherment written between the lines or in the margin, clear passages in the same letter, a deciphered copy in the same file.
     - external key matched: An existing key that had not been connected with this document (a key for another correspondent or series, a key sheet filed elsewhere) was found to fit, and applied.
     - known key applied: The key was already identified for this document or its correspondence (on its archive or DECODE record, published for it, filed with it) and was applied. The work is transcription and decipherment, which is not mechanical when the key is polyphonic or variable-length.
     - existing decipherment: A complete decipherment already existed (between the lines, on a separate copy, or in print). The work was transcribing the cipher and checking the decipherment against it.
     - not solved: Attempted; neither the key nor the text was obtained. What was learned is recorded. A target whose key was obtained but whose text does not read keeps the method that obtained the key, with extent "none (text not read)".
     - not applicable: Not a cipher (clear text, shorthand, a key sheet, instructions), or the item could not be reached.

REPORT BACK
With a clone: open a pull request with your targets/<writer><year>/ folder. Say in the description how you want to be
  credited (name, initials or handle, and a link) and include the same points as the issue below.
Without one: open an issue: https://github.com/dbourdeau/cyphersolver/issues/new?template=solution.yml&title=Solution%3A+catalogue+2646&target=catalogue+2646
(other kinds of contribution: https://github.com/dbourdeau/cyphersolver/issues/new/choose). Include:
  - Target: catalogue 2646
  - the documents and the images you worked from;
  - the key: sign-to-value table, code groups with values, nulls;
  - the decipherment line by line against the ciphertext, with uncertain passages marked;
  - the method (step 7), the share of tokens that give sense, complete or partial (step 6), and whether the
    reading was already published anywhere;
  - your step log (WORKLOG.md), and the tools and the AI model you used.
If you cannot open issues or pull requests, write this report for your human to send. Credit goes to the person who sends it.