Answers are grounded in this paper and cite their sources.
The D’Agapeyeff Cipher (1939): grille recovery and polyphonic-table identifiability under n-gram scoring Research preprint. The report has been through internal verification and AI-assisted critical review. It has not been peer-reviewed. In 1939 Alexander D’Agapeyeff ended Codes and Ciphers with a 196-symbol challenge cipher that has never been read. Prior work (Hendrick 2026) showed that no one-to-one letter key with any transposition can produce its symbol counts. This study tests one construction that fits every observed property: a 14 × 14 turning grille over a polyphonic table, in which each of 13 symbols stands for a group of letters. The question is whether, under that hypothesis, the cipher’s key could be found at all. All rates below were measured on practice ciphers built in the hypothesized form with known keys. What this study adds, within the literature reviewed: Grille recovery when the polyphonic table is known (extends Hendrick’s one-to-one result). Six-symbol scoring recovered at least 45 of 49 grille orbits in 6 of 6 practice ciphers, 5 of them exactly. This is a preliminary demonstration, not recovery of the historical grille.The table must be essentially exact (extends Hendrick’s result to polyphonic tables). With 6 or more letters in the wrong group, the grille was found in 0 of 12 practice ciphers.182 symbols rarely single out the true table under the n-gram scores tested. The search started at the true table and was given the true symbol order. It kept the truth in only 11% of 200 practice ciphers, and in 8–26% across ten alternative setups (2,000 runs: natural prose, three corpora, four scores). The higher-scoring tables fit the particular passage, not English. Keeping the truth passes one in two at roughly 420 symbols.Four-square and two-square with random keys are excluded by symbol usage: none of 10,000 practice ciphers used 13 or fewer symbols.A four-square whose key is built around the message can reproduce the cipher’s usage and flatness: 13 of 100 passages; 10 of 100 were also not flagged by the order test used. This is a feasibility result and an open lead, not evidence that D’Agapeyeff used it.Under the methods tested, the hypothesis can be neither confirmed nor refuted from the ciphertext alone. No decipherment is claimed. These are measurements under stated methods, not a proof that no method could succeed. The work was an AI-assisted cryptanalysis study conceived and directed by Nicolas Houghtaling. Software, analyses and writing were done by Anthropic’s Claude under the author’s direction, supported by a team of AI agents. Drafts were reviewed by ChatGPT (OpenAI); that is an AI review, not peer review. Every result is recorded in a case file with its evidence. Before publication, every published experiment was rerun at full size from a fresh clone, with no network access. Of 4,959 records compared, 4,956 matched, and every printed number was recomputed. That was an internal reproduction by the project, not independent replication. Code, data, evidence and the validation record: https://github.com/voodoomods/dagapeyeff-cryptanalysis-1 Report and data: CC BY 4.0. Code: Apache-2.0. Files derived from wordfreq data: CC BY-SA 4.0.
No takes yet. Share an insight, caveat, or question.
Nicolas Houghtaling (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: