Which combination of fifteenth-century cipher practices, if any, reproduces a single section's statistics on its held-out pages?
One answer, none aheadNone is ahead yet. See what would settle it.
Where it stands
One answer has been proposed. None is ahead on the evidence yet. One is still possible.
Why it matters
Named ciphers tested one at a time each miss several statistics (T15, M-T36, M-T38). Fitting the practices of the period as one family (homophones and how they are chosen, syllable signs, a nomenclator, scribal abbreviation, nulls, spacing practice, line layout) to one section at a time says which features a mechanism needs, and gives a crack attempt a model to aim at. Sections are used so that page order and changes of scribe or subject cannot matter.
Answers 1
Next steps 5
1
Give the T39 signs a word-internal grammar (slot-position sign tables learned from the training pages) and refit: the remaining misses (glyph predictability, one-off words, edge dependency) all say the manuscript's words are more rigidly built than freely joined signs.
~6 h
—not judged
2
Homophone rung (R2 of T44): the search now sits 64 nats from the true key yet 30 % of sign uses are wrong and coverage is flat at 0.47 (bar 0.487). Before more search, diagnose which element pairs the wrong rows confuse (confusion matrix of the best state against the planted key) and whether a richer Latin model (4-gram, or word-level dictionary scoring inside EM) separates them; then rerun R2 with the same effort and the trigram-objective hopping started earlier.
~6 h
—not judged
3
Segmentation-aware solver for the stream family: a segmental hidden Markov model (signs of one to three glyphs as emissions, the sticky homophone column as a latent state) fitted jointly with the key, validated on rung B of M-T40 before any use on the manuscript.
~16 h
—not judged
4
Sharper null for decodes: run the fitted generator on shuffled Latin words, which keeps the cipher's structure (edge dependency, line layer) and removes plaintext order; a reading must beat that, not only the shuffled manuscript.
~3 h
—not judged
5
Picture anchors: the fit favours a nomenclator, so plant names would be whole-word signs; with the plant-identification table (Q-034), test whether page-first words on pages showing the same plant agree, and use agreed identifications as cribs for the key search.
~12 h
—not judged
Ranked by how likely each is to move the key question, for the time it takes
History
- Greshko: period practice as a re-spaced stream gives the closest held-out fit to the herbal pages yetFor: Latin as a re-spaced homophonic stream (period practices)
- Lasry et al.: period practice as a re-spaced stream gives the closest held-out fit to the herbal pages yetFor: Latin as a re-spaced homophonic stream (period practices)
- Session 11: raised the question, proposed one answer, ran five tests and found one result bearing on its answers
- Session 12: took the question up again, proposed one answer and ran 15 tests
- The lattice EM at GPU scale up a ladder of planted controls; the simplest control read, the homophone rung not: took the question up again, proposed one answer and ran one test