Kryptos / K4We asked ChatGPT to work on Kryptos. It ruled out entire sets of keys, checked published claims, and left a reproducible trail. We did not recover the encryption method.
Download the research ↓22–34 EAST NORTHEAST
64–74 BERLIN CLOCK
Original ciphertext. Positions are counted from 1.
24 letters of plaintext are publicly confirmed.
The K3-derived test alone excludes 4,524,029,526,974,464 parameter combinations. A contradiction between two confirmed letters can eliminate every key in a model without trying the keys one at a time.
4.52 quadrillion
Key-and-method combinations excluded by four K3-derived rearrangements, with every 8- or 10-letter key.
4 routes × 2 alphabets × 2 signs × 2 operation orders × (268 + 2610)
Independent verification ↗204,913,783,045,357,568 further labelled assignments are covered by the binary 7 × 7 matrix proof: 364 alphabet/alignment cases, each allowing all 249 binary matrices.
The earlier columnar searches cover larger spaces still. Their complete finite accounting is below. Models can overlap, so these counts describe parameter assignments, not distinct plaintexts or independent solutions.
The sum across the ten families in this ledger is:
656,088,076,618,908,824,313,107,197,286,544,515,861,848,336,008,633,625,137,803,385,820,549,666,122,804,211,668,074,668,032
Model-labelled parameter assignments excluded. Every row retains its source summary. Equivalent keys or cipher functions can appear in more than one model; the total has not been deduplicated across families.
| Family | Negative model cells | Parameter assignments |
|---|---|---|
| Ordinary columnar | 338 | 518,454,579,766,154,252,339,121,208,833,120,690,250,873,497,093,510,325,137,732,846,885,467,299,656,221,982,720,000 |
| Autokey columnar | 432 | 1,529,022,446,773,476,595,988,542,495,064,064,000 |
| Independent column reversal | 44 | 4,313,644,366,707,057,999,696,887,808,000 |
| Mixed alphabet axes | 20 | 935,514,845,880,975,360,000 |
| Progressive keys | 48 | 1,872,361,385,492,742,144,000 |
| Column-reset roots | 4 | 11,702,924,506,195,034,112,000 |
| Affine substitution | 20 | 3,329,234,326,978,560,000 |
| Inverse columnar, including derived periods | 56 | 656,087,558,164,329,058,158,854,858,165,335,682,741,158,085,135,136,531,625,949,221,327,284,922,838,118,603,545,837,568,000 |
| K3-derived routes | 64 | 4,524,029,526,974,464 |
| Binary 7 × 7 matrices | 364 | 204,913,783,045,357,568 |
The count omits repeated runs, the surviving width-31 model, baseline replications, and the unbounded key-length extension of the reset proof. The Berlin ring experiment is reported separately: 2,212 registered cases fail under the distinct-hour-number interpretation.
Download formulas, scope and source ledger ↗Six completed goals used 8,403,684 tokens. The interrupted initial run had already used 366,344 at its last recorded checkpoint. The combined figure is therefore a documented lower bound.
| Research phase | Recorded tokens |
|---|---|
| Initial run, last recorded checkpoint | 366,344+ |
| Main cryptanalytic campaign | 3,417,704 |
| Public-source survey | 1,310,021 |
| Four follow-up investigations | 1,128,952 |
| Binary-matrix and archive continuation | 1,233,387 |
| Berlin World Clock investigation | 720,361 |
| K3-derived alternatives | 593,259 |
These are native goal usage records, not token budgets or a billing estimate. Planning, unbudgeted follow-up, and this website are excluded. The old 1.31M figure counted only the public-source survey.
| Family | Bounded scope | Outcome |
|---|---|---|
| Ordinary periodic columnar | Widths 16 to 20, periods 1 to 6; width 18, periods 7 to 12; AZ and KRYPTOS alphabets, both signs and stages | closed |
| Ordinary width 31 | Periods 1 to 5 under the same conventions | 36 models closed |
| Width 31, period 6 | One additive-before model with key XETQGR | structural fit only |
| Ciphertext autokey | Widths 16 to 20, lags 1 to 12, two alphabets and three signed conventions | 360 models closed |
| Plaintext autokey | Width 18, lags 1 to 12, same conventions | 72 models closed |
| Mixed axes, progressive, reset, affine | Declared fixed schedules and finite parameter cells documented in the report | bounded negatives |
| Inverse columnar | Source and destination phase models with local certificates, plus exact source-phase period 4 closure | declared cells closed |
A negative result applies only to the stated alphabet, period, route, operation order, and clue alignment. It does not close the named cipher family.
Each finding applies to a stated model or implementation. The linked reports preserve the equations, source references, and verification.
A fixed binary matrix model proposed in Richard Bean's 2018 notes was tested against the newer EAST, NORTHEAST, BERLIN, and CLOCK plaintext. All 364 registered alphabet and alignment cases were rejected, including the 99 cases that survived the first partial gate.
A separate public proposal displayed a 97-position shift array beside its ciphertext and candidate plaintext. Direct A=0 arithmetic disagreed at 75 of 97 positions. Two independent implementations matched.
Inspection found transformed clue constraints applied incorrectly and a ragged-grid mapping that dropped real cells while admitting invalid indices. A planted counterexample was rejected by the published formulas.
A literal fitted counterexample refuted the blanket extension of direct-alignment Bean constraints through arbitrary permutations. The counterexample is a proof defect, not a K4 candidate.
The width-31, period-6 model produced a witness using the key XETQGR. It reproduces all 97 ciphertext letters and matches all 24 disclosed plaintext letters.
The remaining 73 letters are incoherent. Nine clue-free columns can also be rearranged in 362,880 ways without breaking the disclosed clues. There is no source-derived reason to prefer this witness, so it remains a structural fit and nothing more.
BERLINCLOCK refers to the Urania World Clock at Alexanderplatz. The shaft map and the drum have a documented vertical city-to-time-zone relationship.
No acquired plan fixes the pre-1990 city-sector order to geographic north. Modern labels are unsafe because a 1997 restoration added 20 city names and changed time-zone treatment.
Acquire a dated plate schedule or north-referenced drawing before testing another period-24 city stream. Unknown historical cells must remain unknown.
candidate gate
Require an exact 97-character round trip, all 24 disclosed letters, coherent text outside the clues, an independently motivated route and key, and independent verification. A necessary survivor is not a plaintext candidate.
Kryptos was installed at CIA headquarters in 1990. Its fourth passage contains the 97 letters shown above.
Plaintext was found in archival material in 2025. That discovery is separate from independently reconstructing the encryption method, which this campaign did not achieve.
Read the archival discovery report ↗