
How to read the archive§
Each card identifies the kind of evidence, links to a source, and states what would have to be shown before a proposed reading should be treated as established. That matters especially for tiny cryptograms, where a name or phrase can be fitted after the fact. For the active Kryptos challenges, see the separate Kryptos research explorer.
D’Agapeyeff Cipher§
1939
Alexander D’Agapeyeff published this 395-digit exercise in the first edition of Codes and Ciphers. No plaintext or method has been verified; a construction or transcription error remains possible.
What holds up
- Cut as consecutive pairs, all 395 digits fall into place: the first digit of every pair lies in {6,7,8,9,0}, the second in {0,1,2,3,4,5} — zero exceptions, leaving 197 pairs and one trailing
0. Cutting one digit later scores 390 violations, so the split is real and positioned. - It is not the digit mix talking. Of 2000 shuffles of the same digits, none matched the alignment; treating the two seats as independent would land roughly 202 of 395 by chance.
- So the text is a 5×6 coordinate grid in which only 19 of the 30 cells are ever visited —
01 02 03 05 60 61 70 73 80 90 95stay empty. A 5×5 Polybius square with one row and one column renumbered cannot produce a six-value second digit, which is what kills the most-cited explanation. The one wrinkle:0serves both seats (04and90), so the seats overlap by exactly that digit. - The pairs are not a normal alphabet:
81×20, then75 62 82 85×17,64×16,83×15,74×14,91 63×12,84 65×11,72×9, then a cliff to92×3,93×2 and04 94 71 00×1. Thirteen codes carry 95% of the message at 8–10% each. An earlier draft of this card read that flatness as a mark against English, and it is not one. Scored freely against the English profile the cipher sits at chi-square29.6; noise drawn from the English profile scores~0and noise drawn from the cipher's own profile scores29.6, so it matches the second null exactly — but 382 letters of genuine English score19.2, sitting between the two. A test on which real English does not cleanly beat the cipher cannot convict the cipher, andprofile_fitis checked in only to record that trap. - The one-cell-per-letter reading can now be measured rather than argued. Attacking the 197 pairs as a keyed grid scores
-5.55log₁₀/char under an English quadgram model. The ceiling — real English of the same length, cut to 19 letters, attacked identically — reaches-4.39, and profile-matched nulls (random streams carrying the cipher's own flat frequency shape) fall between-5.26and-5.87. The cipher lands inside the null, not near the ceiling: the attack is strong enough to solve a genuine grid at this length and finds nothing. - The stronger refutation needs no attack at all, because it does not depend on which letter any cell means. Two coincidence statistics — how often adjacent codes repeat relative to a shuffle of the same frequencies, and how much digram structure exists beyond what the singles already explain — give
1.264and0.064for the cipher, against0.502and0.136for English. That is 4.5σ and 2.7σ from English and about 1σ from a plain shuffle. Because relabelling cannot move either number, and a transposition could only push them toward English, this closes the keyed grid with a columnar or route layer on top — the usual fallback when the plain reading fails. - Two families fall to counting, not testing. Playfair, two-square, four-square and bifid-digraph cannot emit the same plaintext letter twice in a row, and a keyed grid maps each letter to one fixed code, so such a cipher would show zero adjacent repeats whatever the key — the text shows 17 (thirteen pairs, two runs of three). And a per-digit rule over a decimal stream has ten inputs, so its image holds at most ten distinct output pairs: the running-constant reading — digits of π or e, at any stride, under any renumbering — cannot produce the cipher's nineteen. Neither argument involves a source distribution, a smoothing prior or a key.
- The strongest measurement is out-of-sample, so nothing is free to overfit: leave-one-out, a Markov model beats an iid one by
-47nats on the cipher, against-42 ± 6for a shuffle of the same multiset (z = −0.7) and+13 ± 8for 197 letters of English — seven sigma away. The gap holds at every smoothing prior tried (α = 2, 9.5, 50), and the test was validated against a chain that genuinely carries memory. Read left to right, the preceding pair predicts the next no better than a bag of the same symbols does. - The one statistic leaning the other way: a Wald–Wolfowitz run test on the 13 busiest codes puts the cipher
1.8shuffle-sds more clustered than randomness, p ≈ 0.07, with burstiness8.85against8.17for independent draws and5.88for English (eight corpus slices — thin, and stable only to ±0.07). It is correlated with the repeat count rather than a fourth opinion, so it is a clue, not a result. - One detail supports the terminator reading of the odd cell:
00appears exactly once, as the final pair. Strip it and the second seat is the clean five-value set{1,2,3,4,5}, leaving 18 content codes — a 5×5 with seven unused cells, plus one stray0past the end that no pair completes. - That leaves the grid meaning something other than one letter per cell — a digraph or word codebook, a second layer beneath the coordinates, or a damaged 1939 printing — and each of those survives a negative test that would have convicted the first.
- Reproduce with
python3 scripts/dagapeyeff_alignment.py(structure and null test) andpython3 scripts/dagapeyeff_detect.py --full(ceiling, null, attack) and--mapfree(the mapping-invariant tests, the run test and the two counting arguments). Frequency-shape arguments are deliberately absent from that list:profile_fitis a library function, not a mode, because the four numbers above are all it produces and none of them separate the cipher from noise. A solver will still report a confident key for this text — that is the failure mode the boundary below exists to catch.
Verification boundary: An English-looking fragment is not a solution without an exact, reproducible construction for the published digits.
Zodiac Z13§
1970
This short cipher follows the phrase “My name is—” in a Zodiac letter. Its length leaves too little information to distinguish a proposed name from other fitting candidates, and no definitive solution is accepted.
Verification boundary: Treat proposed names as hypotheses, not identifications, unless a method supplies unique and independently checkable validation.
Zodiac Z32§
1970
This 32-symbol message accompanied a San Francisco Bay Area map and was said to concern a bomb location. It has not been definitively decoded; map context does not by itself turn a candidate reading into proof.
Verification boundary: A credible solution must account for the full symbol sequence and produce a falsifiable connection to the accompanying map.
Dorabella Cipher§
1897
Composer Edward Elgar’s note to Dora Penny uses a small set of curved glyphs. Textual and musical interpretations have been proposed, but no reading has gained consensus as the intended message.
Verification boundary: The short text admits many plausible readings; historical fit and an exact, consistently applied key are required before calling one a solution.
Beale Ciphers 1 & 3§
Published 1885
Of the three number ciphers in the Beale Papers, only cipher 2 has a demonstrated Declaration of Independence key. The location and heir messages, ciphers 1 and 3, remain open—and the underlying treasure story itself is unverified.
Verification boundary: The archive separates the unsolved texts from the historical claim: a decryption would not by itself authenticate the treasure narrative.
Voynich Manuscript§
15th–16th century
Beinecke MS 408 is written in an unidentified script by an unknown author. It is not established that the writing is a cipher at all; cryptographic approaches have not produced an accepted decipherment.
Verification boundary: This belongs in an undeciphered-script archive, not in a list of solved cipher systems. Yale provides high-resolution research images.
Phaistos Disc§
Bronze Age Crete
The unique spiral object bears more than 240 stamped signs. Its script, language, purpose, and even whether it should be treated as ordinary text remain unresolved because there is no comparable corpus or bilingual key.
Verification boundary: Proposed readings are not established decipherments. The archive labels it an undeciphered object rather than assuming a specific cipher mechanism.
What is not a solution§
A readable fragment, a favorable n-gram score, a thematic story, or a coincidence with a map or historical event is evidence to investigate—not a decipherment. A strong solution gives a complete method, applies it consistently to the source, and leaves a result that another researcher can reproduce.
Frequently asked questions§
Does unsolved mean nobody has proposed an answer?
No. Many of these objects have many proposed readings. Here, unsolved means there is no generally accepted, reproducible decipherment that explains the full source material without arbitrary choices.
Are all of these conventional ciphers?
No. The D’Agapeyeff and Zodiac entries are ciphertext challenges; the Voynich Manuscript and Phaistos Disc are undeciphered written objects. The distinction is stated on every card because it changes what a valid solution would look like.
Why are short cryptograms especially difficult?
Short text provides too little redundancy to choose uniquely among keys and plaintexts. A candidate can look meaningful while many different candidates fit equally well. Independent verification is essential.